title
stringlengths 10
190
| abstract
stringlengths 118
1.9k
| related work
listlengths 2
8
| references
listlengths 0
0
|
|---|---|---|---|
An Efficient Genetic Algorithm for Discovering Diverse-Frequent Patterns
|
Working with exhaustive search on large dataset is infeasible for severalreasons. Recently, developed techniques that made pattern set mining feasibleby a general solver with long execution time that supports heuristic search andare limited to small datasets only. In this paper, we investigate an approachwhich aims to find diverse set of patterns using genetic algorithm to minediverse frequent patterns. We propose a fast heuristic search algorithm thatoutperforms state-of-the-art methods on a standard set of benchmarks andcapable to produce satisfactory results within a short period of time. Ourproposed algorithm uses a relative encoding scheme for the patterns and aneffective twin removal technique to ensure diversity throughout the search.
|
[
"Many variants of pattern set mining are investigated in the literature. Among them to find patterns which are correlated {{cite:f27c96151f8c0ebdbd80cf0c9ead0d31e1b04031}}, discriminative {{cite:e8041a08e3be08841a83194ba88dc2e6b8d4d8b3}}, contrast {{cite:4a82cfda80088c4f25d7a3e0682ca91ea297c83a}} and diverse {{cite:660dd07b6795fe25b6640762822393318574c169}} became promising tasks.Various algorithms has been proposed as a general framework for pattern mining {{cite:10e11b04f9ce3b0068889b02260127765d907d73}}, {{cite:d5e8f44128b5c846a3d4c1d23691c3f0ebfd999f}} in last few years. Many languages have been developed for declaratively modeling problems, such as Zinc {{cite:e6c2f7f1e7b68688b8fe99f8eb2cd0eb56f88625}}, Essence {{cite:670cd555f57ed27b459ce5db9e94ae9b3fc2a329}}, Gecode {{cite:e3d1a305c51942210d14e8eaf8cbf384eb950869}} and Comet {{cite:10e11b04f9ce3b0068889b02260127765d907d73}}, {{cite:b558f76f50940a6e8f941a99ceb8acf50f7f5026}}.",
"To search and prune the solution space, most of these methods use systematic search methods and the algorithms, those are not only exhaustive in nature but also take huge amount of time. On the other hand, stochastic search algorithms does not guarantee optimality but give a approximately best results within a short period of time. However, Gunset al. {{cite:10e11b04f9ce3b0068889b02260127765d907d73}} investigated a technique by simplifying pattern set mining tasks and search strategies by putting these into a common declarative framework. In a recent work, Hossain et al. {{cite:fad0a859db0c2ea2e1e96aabfb881a85e44ea158}} explored the use of genetic algorithms and other stochastic local search algorithms to solve the concept learning task using small datasets."
] |
[] |
A coalgebraic semantics for causality in Petri nets
|
In this paper we revisit some pioneering efforts to equip Petri nets withcompact operational models for expressing causality. The models we propose havea bisimilarity relation and a minimal representative for each equivalenceclass, and they can be fully explained as coalgebras on a presheaf category onan index category of partial orders. First, we provide a set-theoretic model inthe form of a a causal case graph, that is a labeled transition system wherestates and transitions represent markings and firings of the net, respectively,and are equipped with causal information. Most importantly, each state has aposet representing causal dependencies among past events. Our first resultshows the correspondence with behavior structure semantics as proposed byTrakhtenbrot and Rabinovich. Causal case graphs may be infinitely-branching andhave infinitely many states, but we show how they can be refined to get anequivalent finitely-branching model. In it, states are equipped withsymmetries, which are essential for the existence of a minimal, oftenfinite-state, model. The next step is constructing a coalgebraic model. Weexploit the fact that events can be represented as names, and event generationas name generation. Thus we can apply the Fiore-Turi framework: we model causalrelations as a suitable category of posets with action labels, and generationof new events with causal dependencies as an endofunctor on this category. Thenwe define a well-behaved category of coalgebras. Our coalgebraic model is stillinfinite-state, but we exploit the equivalence between coalgebras over a classof presheaves and History Dependent automata to derive a compactrepresentation, which is equivalent to our set-theoretical compact model.Remarkably, state reduction is automatically performed along the equivalence.
|
[
"This paper follows a line of research on coalgebraic models of causality, started in {{cite:b2f53c5a28826b1babcf835afdecd1ec5df1ed3c}} by the same authors. The categorical machinery is the same in both papers, namely presheaf-based coalgebras, HD-automata, and the equivalence among them. However, this paper takes a further step towards a general categorical theory of causality. In {{cite:b2f53c5a28826b1babcf835afdecd1ec5df1ed3c}}, in fact, we have provided models for a particular class of causal LTSs, namely Degano-Darondeau ones. In this paper, instead, we treat Petri nets, which are much more general. For instance, unlike Degano-Darondeau LTSs, Petri nets can describe synchronizations of more than two processes.",
"In {{cite:b2f53c5a28826b1babcf835afdecd1ec5df1ed3c}} we start from existing set-theoretic models, similar to abstract CGs, whereas the models we introduce here are novel. In both papers we represent causal dependencies as posets over events, but in {{cite:b2f53c5a28826b1babcf835afdecd1ec5df1ed3c}} events are unlabeled and are canonically represented as natural numbers. Here we have labels and we take a more general approach: instead of choosing specific representatives of events, we make abstract CGs parametric in this choice. This requires more technical work and it further validates the categorical approach, where book-keeping details are abstracted away. The categorical environment in this paper is more elaborate than {{cite:b2f53c5a28826b1babcf835afdecd1ec5df1ed3c}}, due to labeling. In particular, event generation is more complex, and is studied in greater detail. Another difference is that here we give conditions under which the model with only immediate causes is finite, whereas in {{cite:b2f53c5a28826b1babcf835afdecd1ec5df1ed3c}} decidability is not treated.",
"A first version of HD-automata for Petri nets, called causal automata, has been introduced in {{cite:37ea9156ec14b7b9576882b8368e5f47f6fa663c}}. However, their construction is purely set-theoretical and does not include symmetries, so the existence of a minimal model is not guaranteed. This version of HD-automata is similar to what we call immediate causes CG (without symmetries). HD-automata with symmetries were developed for the {{formula:172b2af3-6e02-4966-a16a-eb61c440cab5}} -calculus in {{cite:cd2428b24ab6e51e44de9da2b1dd051b2c86d54b}}, {{cite:d70c881c14009b612c76002d8104f010dd36116a}}, and a general categorical treatment was provided in {{cite:970c2a979ed2bededcedb383daeb876af1f64a42}}. In all these cases nominal structures associated to states are just a sets of (event) names, whereas we have posets, which are more adequate to represent causal dependencies.",
"We can cite {{cite:74d2835b3dbbf7a4e90d8df51798741c0e7468cc}} for the introduction of transitions systems for causality whose states are elements of presheaves, intended to model the causal semantics of the {{formula:6db82432-7bbd-4731-80a2-a2e3d8b22516}} -calculus as defined in {{cite:c926fc632c305a7950843a6e8d3854e15e3b8935}}. However, the index of a state is a set of names, without any information about events and causal relations. The advantage of our index category is that it allows reducing the state-space in an automatic way, exploiting a standard categorical construction. This cannot be done in the framework of {{cite:74d2835b3dbbf7a4e90d8df51798741c0e7468cc}}.Finally, an HD-automaton for causality has been described in {{cite:970c2a979ed2bededcedb383daeb876af1f64a42}}, but it is derived as a direct translation of causal automata and its states do not take into account causal relations.",
"Other related works are {{cite:b0e1871d627a839b7e18a3f9ea760c9d08f7a89e}}, {{cite:89e3fe7622badc38b66cbe398b81f311a2045495}}, where event structures have been characterized as (contravariant) presheaves on posets. While the meaning of presheaves is similar, the context is different: we consider the more concrete realm of coalgebras and nominal automata. A more precise correspondence with such models should be worked out."
] |
[] |
VMF-SNE: Embedding for Spherical Data
|
T-SNE is a well-known approach to embedding high-dimensional data and hasbeen widely used in data visualization. The basic assumption of t-SNE is thatthe data are non-constrained in the Euclidean space and the local proximity canbe modelled by Gaussian distributions. This assumption does not hold for a widerange of data types in practical applications, for instance spherical data forwhich the local proximity is better modelled by the von Mises-Fisher (vMF)distribution instead of the Gaussian. This paper presents a vMF-SNE embeddingalgorithm to embed spherical data. An iterative process is derived to producean efficient embedding. The results on a simulation data set demonstrated thatvMF-SNE produces better embeddings than t-SNE for spherical data.
|
[
"This work belongs to the extensively studied area of dimension reduction and data embedding. Most of the related work in this field has been mentioned in the last section. Particularly, our work is motivated by t-SNE {{cite:3e978b6be7f494b088c4f6b3c8f34d41eba065fd}}, and is designed specifically to embed spherical data which are not suitable to be processed by t-SNE. A more related work is the parametric embedding (PE) {{cite:fbffa5f32ac9eda704070d63ca4981a2a6d60f4f}}, which embeds vectors of posterior probabilities, thus sharing a similar goal as our proposal: both attempt to embed data in a constrained space though the constrains are different ({{formula:6e1ed68b-35e2-48fc-a1aa-6e036dcdfa7e}} -1 in PE and {{formula:def56698-3e71-4487-b99d-eb3954de01df}} -2 in vMF-SNE).",
"Probably the most relevant work is the spherical semantic embedding (SSE) {{cite:369e7853c5cd172d31e1c4fb17c75c2e0c7b90c0}}. In the SSE approach, document vectors and topic vectorsare constrained on a unit sphere and are assumed to follow the vMF distribution. The topic model and the embedding model are then jointly optimized in agenerative model framework by maximum likelihood. However, SSE infers local similaritiesbetween data samples (document vectors in {{cite:369e7853c5cd172d31e1c4fb17c75c2e0c7b90c0}})using a pre-defined latent structure (topic vectors), which is difficult to be generalized to other tasks as thelatent structure in most scenarios is not available. Additionally,the cost function of SSE is the likelihood, while vMF-SNE uses the symmetric KL divergence."
] |
[] |
Mobile Multi-View Object Image Search
|
High user interaction capability of mobile devices can help improve theaccuracy of mobile visual search systems. At query time, it is possible tocapture multiple views of an object from different viewing angles and atdifferent scales with the mobile device camera to obtain richer informationabout the object compared to a single view and hence return more accurateresults. Motivated by this, we developed a mobile multi-view object imagesearch system, using a client-server architecture. Multi-view images of objectsacquired by the mobile clients are processed and local features are sent to theserver, which combines the query image representations with early/late fusionmethods based on bag-of-visual-words and sends back the query results. Weperformed a comprehensive analysis of early and late fusion approaches usingvarious similarity functions, on an existing single view and a new multi-viewobject image database. The experimental results show that multi-view searchprovides significantly better retrieval accuracy compared to single viewsearch.
|
[
"Due to the recent developments in mobile devices with cameras, there has been a growing interest in mobile visual search, and research works investigate different aspects of it, such as architectures, power efficiency, speed, and user interaction. Chen and Girod {{cite:a8a8c0f83f9fea604a928104a953bb9de573e555}} describe a mobile product recognition system where the products are CDs, DVDs and books that have printed labels. The system is local feature based, and Compressed Histogram of Gradients (CHOG) and Scale-Invariant Feature Transform (SIFT) are used as local features. Two client-server architectures are implemented and compared in terms of response time: one is sending images, the other one is extracting features on the client and sending the features. Sending features took five seconds, sending images took ten seconds to respond. This means that over slow connections like 3G it is faster to extract and send features.",
"Storage space and retrieval speed are important in mobile visual search. Girod et al. {{cite:4cd1d83abf1e40aad3d35bd94588a7fd1b4fd867}} describe a mobile visual search system that adopts the client-server architecture in which the database is stored on the phone. The system uses the Bag-of-Words (BoW) approach, and 4 different compact database methods are experimented and their performances are compared. Li et al. {{cite:f35f02f2741a307a13eadc47bd30ca0b214ad87e}} propose an on-device mobile visual search system. The system uses the BoW approach with a small visual dictionary due to the memory limitation. Additionally, the most useful visual words are selected to decrease the retrieval time considering the processor limitation. Guan et al. {{cite:d3bfad4f07b4c3d6d75a71ddcf7cf875888401f7}} describe an on-device mobile image search system, which is based on bag-of-features (BoF) approach. The system uses approximate nearest neighbor search to use high dimensional BoF descriptors on the mobile device with less memory usage. The search system also utilizes the GPS data from the mobile device to reduce the number of images to be compared. Zhou et al. {{cite:a149e4a571aac45c6843f6839392352c4d209647}} describe a mobile image search system that does not use a codebook, and hence, decreases the memory usage and quantization errors. The system makes use of SIFT descriptors, and the dimensionality of the descriptors is reduced using Principal Component Analysis (PCA).",
"Mobile devices have high user interaction potential; this has been utilized for better retrieval.Joint Search with Image Speech and Words (JIGSAW) {{cite:9886232f3bdf455f0169f3677b4806623374a44d}} is a mobile visual search system that provides multimodal queries. This system allows user to speak a sentence and performs text-based image search. The user selects one or more images from the result set to construct a visual query for content-based image search. In {{cite:a975c7e4e46e2a504dd2de145b91e278dc47598f}}, a mobile product image search system that automatically extracts the object in the query image is proposed. From the top n images that have a clean background, object masks are found. The object in the query image is then extracted by using a weighted mask approach and its background is cleaned. The cleaned query image is finally used to perform image search. Extracting the object-of-interest and performing the query with a clean background is shown to work better. Similarly, TapTell {{cite:fc0074d8a11f9f19f5f6988dc1ce456bd75fef56}} is an interactive mobile visual search system, in which users take a photo and indicate an object-of-interest within the photo using various drawing patterns.",
"There are several mobile image search applications available on the mobile market.Point&Find {{cite:4f2847a3e8370720b1724acb280cd46ac0ce7714}} allows the users point the camera to the scene or object and get information about it. Kooaba is a domain-specific search system whose target domains are books, DVD and game covers {{cite:5d6541f2d38479e84b155f9b41cb6f49dbdad08f}}. Digimarc Discover {{cite:de955a6b3705269105e41d49b3eabf6ecb20621f}} is similar to Point&Find; the user points the camera to an object and gets information about it. PlinkArt {{cite:2bafbdf9a05981957c0b7f6aba8ab7432e79d09e}} is another domain-specific mobile search system whose target domain is well-known artworks. The user takes a photo of a well-known artwork and gets information about it. One of the latest mobile search application is CamFind {{cite:5adf3b1dab0156269a1b3413eddd2952931f6c55}}, which is a general object search system. When the user takes a photo of a scene, products are identified and similar objects are listed as a result. Another recent mobile search application is Amazon Flow {{cite:bc806be2a585c10a907fc985260cf68d96594ac0}}; the user points the camera to the object and receives information about it.",
"Multi-image queries have been used to improve image retrieval. Arandjelovic and Zisserman {{cite:5ff01bad2cb90e174b8725e1141312b6291c76f7}} propose an object retrieval system using multiple image queries. The user enters a textual query and Google image search is performed using this textual query. The top eight images are then retrieved and used as query images. Early and late fusion methods are applied. Tang and Acton {{cite:29d492210f63c6e937cc623636fe1cbb0060cc50}} propose a system that extracts different features from different query images. These extracted features are then combined and used as the features of the final query image. Another system is proposed in {{cite:aa1bccdd3541ed6a98bb032184487c4a11daf21f}}, which allows users to select different regions of interest in the image. Then each region is treated as queries and results are combined. Zhang et al. {{cite:3f57772c1787d563a36e58a59f18848891ac41fb}} describe a similar system, which also uses regions; however, these regions are extracted automatically and users select regions from the extracted parts. Xue et al. {{cite:07b73138369ae9e60e1b9f15f2a4c06e4f667507}} propose a system that uses multiple image queries to reduce the distracting features by using a hierarchical vocabulary tree. The system focuses on the parts that are common in all the query images. The multi query system described in {{cite:c8f00aa8dad3c1404cc76696a39321ccdc3f59f3}} uses early fusion; each database image is compared with each query image and each query image gets a weight according to the similarity between the query image and the database image.",
"Queries using multiple images have been utilized in several existing works to improve the performance of visual search systems. In this work, we take this one step further and leverage the interaction capability of mobile devices to perform more informative visual object queries by taking multi-view images of the query object. We performed comprehensive analysis of various similarity functions and early/late fusion methods in terms of retrieval precision and running time. Furthermore, we collected a new multi-view object image dataset to perform single and multi-view query experiments."
] |
[] |
Estimation of the Diffusion Constant from Intermittent Trajectories with
Variable Position Uncertainties
|
The movement of a particle described by Brownian motion is quantified by asingle parameter, $D$, the diffusion constant. The estimation of $D$ from adiscrete sequence of noisy observations is a fundamental problem in biologicalsingle particle tracking experiments since it can report on the environmentand/or the state of the particle itself via hydrodynamic radius. Here wepresent a method to estimate $D$ that takes into account several effects thatoccur in practice, that are important for correct estimation of $D$, and thathave hitherto not been combined together for estimation of $D$. These effectsare motion blur from finite integration time of the camera, intermittenttrajectories, and time-dependent localization uncertainty. Our estimationprocedure, a maximum likelihood estimation, follows directly from thelikelihood expression for a discretely observed Brownian trajectory thatexplicitly includes these effects. The manuscript begins with the formulationof the likelihood expression and then presents three methods to find the exactsolution. Each method has its own advantages in either computationalrobustness, theoretical insight, or the estimation of hidden variables. We thencompare our estimator to previously published estimators using a squared logloss function to demonstrate the benefit of including these effects.
|
[
"Historically, one of the primary techniques for estimating the diffusion constant from trajectories relied on a linear regression of the mean-squared-displacement (MSD) of the tracked particle coordinates as a function of time lag {{cite:1f01fa7cbe74317c1f9ff7f1e8700cd011093fc2}}. In the absence of measurement errors, the observed MSD for pure Brownian motion scales linearly with time lag and intersects at the origin, allowing the direct recovery of the diffusion constant from a linear regression on the well sampled data points. It has been shown that a regression of the MSD with an offset parameter can be interpreted to account for the cumulative effects of static {{cite:f94b83c0387b10005b84b2f28a4574fceef69060}} and dynamic measurement errors {{cite:172151358f39889dc03557981a72070ff1b6f725}}. If the MSD is built using the same data points for multiple time lags, the correlation between MSD values must also be taken into account in the regression {{cite:1f01fa7cbe74317c1f9ff7f1e8700cd011093fc2}}, {{cite:94fef8743bb9637e2e1604f82ced70bec4d7d75b}}, {{cite:a40442621b4c7f09c657768b8da528ddb5fd7a1c}}. Although it seems theoretically possible to include individual localization error into the MSD regression, to date this has not been described.",
"A separate line of work has focused on maximum likelihood approaches to the estimation procedure. A maximum likelihood estimator works by finding the maximum of a likelihood function {{formula:e3e784e1-2e95-4755-991c-aa9ae8fad12e}} that gives the probability of observing a particular trajectory {{formula:f9f8c0ee-0f42-4507-885d-9b14f04b72ed}} , given a diffusion constant {{formula:a4b5257a-3331-40af-8cb9-baccd178136f}} . Ideally this probability should incorporate both the variable localization errors of the trajectory and effect of motion-blur. The motion-blur effect arises from the fact that each localization is performed on data that is acquired over some non-zero exposure time. Typically camera sensors integrate the signal over the exposure time resulting in a blurring of the particle image. This blurring has important numerical effects on the likelihood function {{cite:fe7ba68e79caebabb2201fbd559bdccb84bd0fe7}}. A specific solution to the likelihood function has been accurately derived that incorporates the effects of motion-blur but with the caveat that only a single global localization error estimate is used as an input or estimated {{cite:6c0b2b6ce5f24f4211a0cbe44d78781d317fd840}}, {{cite:a40442621b4c7f09c657768b8da528ddb5fd7a1c}}. This estimator is a more robust alternative to the MSD-based estimators because it can implement all trajectory information without incurring systematic error when the data is not well conditioned for a linear regression. Subsequent work has extended this approach to deal with non-uniformly spaced or intermittent trajectories {{cite:627c8d98e5f38de710cefbf17efedbd526073e79}}, however the particular implementation in {{cite:627c8d98e5f38de710cefbf17efedbd526073e79}} did not account for motion blur. Maximum likelihood estimators are not the only class of diffusion estimators that have evolved recently; continued development on displacement-based estimators has resulted in an estimator that incorporates the effects of covariances between sequentially observed displacements {{cite:f31f46804760137fdc57c62dc6ba0a0a894405e1}}.",
"In this work we provide a generalized solution to the likelihood function, incorporating variable localization errors and variable displacement periods, which results in an improvement in estimation accuracy for short trajectories, trajectories with large variations in localization accuracy, and trajectories with intermittently spaced measurements. In Sec. we formulate the diffusion likelihood function to directly incorporate the effects of motion-blur, variable localization errors, and intermittent or non-uniformly spaced observations in time. We present three independent solutions to this likelihood function. The first derivation, the recursive method (Sec. ), is a sequential integration of the nuisance parameters and provides the fastest numerical implementation. The second derivation, the Laplace method (Sec. ), utilizes a second order Taylor expansion to express the likelihood as a multivariate Gaussian in the basis of integration. The Laplace method additionally returns the maximum likelihood values of the true positions given a sampled {{formula:3f18e6a2-37b0-4d88-84d1-3bd12260fc41}} . The third derivation, the Markov method (Sec. ), calculates the characteristic function in order to express the likelihood in the basis of displacements. The Markov method allows us to verify that the generalized form of the expression derived in {{cite:6c0b2b6ce5f24f4211a0cbe44d78781d317fd840}} is the same distribution as the expressions derived in this manuscript. The Markov method was also instrumental in determining the coefficients necessary to reduce the computational complexity of all the methods (Sec. REF ). Each of these derivations leads to an independent, numerically accurate computational algorithm for estimating the likelihood of {{formula:ea91b718-3377-4403-be8d-2bb13927dbc7}} (Sec. ), making full use of all the information contained in a noisy trajectory. The resulting likelihood calculation allows for robust computations in specific problems, such as a maximum likelihood estimator, maximum a posteriori estimate, or change point analysis. We compare the results of our maximum likelihood estimator (MLE) to the current state of the art estimation software {{cite:a40442621b4c7f09c657768b8da528ddb5fd7a1c}} with the squared log loss function and demonstrate that the additional information provided from the localization errors allows for better estimates of {{formula:848f15ba-9384-4401-9b63-a8f437c24bae}} with trajectories parameterized by any non-constant, but otherwise arbitrary distribution of localization variances."
] |
[] |
Bufferless NOC Simulation of Large Multicore System on GPU Hardware
|
Last level cache management and core interconnection network play importantroles in performance and power consumption in multicore system. Large scalechip multicore uses mesh interconnect widely due to scalability and simplicityof the mesh interconnection design. As interconnection network occupiedsignificant area and consumes significant percent of system power, bufferlessnetwork is an appealing alternative design to reduce power consumption andhardware cost. We have designed and implemented a simulator for simulation ofdistributed cache management of large chip multicore where cores are connectedusing bufferless interconnection network. Also, we have redesigned andimplemented the our simulator which is a GPU compatible parallel version of thesame simulator using CUDA programming model. We have simulated target largechip multicore with up to 43,000 cores and achieved up to 25 times speedup onNVIDIA GeForce GTX 690 GPU over serial simulation.
|
[
"Performance of most commonly used shared cache (physically shared) is pathetic in terms of access latency and interference, and also as core counts of chip multicore have direct impact on pressure on the shared on-chip cache. Many researchers proposed technique to reduce interference either by partition efficiently or managing efficiently {{cite:2cff622588ec20493e3780bf49ce333ff7d714d8}}, {{cite:c35a5a29b00f1caf74808803dfb2a56acf78b81a}} to maximized the performance of multicore system.",
"Distributed cache reduces interference and latency, but may not be good where a core requires a larger portion of cache {{cite:bf141d1ac01b874b4efb58f8c8bd58ed4b8387cb}}. And in that case a distributed cache slice may suffer from increased local cache pressure and eviction. So LSPD model provide an effective trade-off between cache access time and share. These papers {{cite:37f495f8466e4dbca030141d06e51ff73736ffe2}}, {{cite:1f06ce3fe20abe5ebe6d981669c96e2cc5f992aa}}, {{cite:5bc93a490b30bf4349662c8ada2d3cba004b8c8b}}, {{cite:ed5e7f08a6727d0a2e666eac595e89def31c4d8c}} describes about various variation of LSPD cache. Also they proposed many logically sharing policy, placement, eviction, replication and management of physically distributed cache to improve the performance of multicore system. In {{cite:5bc93a490b30bf4349662c8ada2d3cba004b8c8b}}, they used a mechanism called cooperative caching, which put a locally evicted block in another on-chip L2 cache that may have spare capacity, rather than evict it from the on-chip hierarchy entirely. In {{cite:ed5e7f08a6727d0a2e666eac595e89def31c4d8c}}, {{cite:37f495f8466e4dbca030141d06e51ff73736ffe2}}, authors proposed architecture with each core having a cache with configurable or adoptable boundary between private and shared portion. In {{cite:37f495f8466e4dbca030141d06e51ff73736ffe2}}, author classified cache access patterns of a range of server and scientific workloads into distinct classes, where each class is amenable to different block placement policies.",
"There are many NoC simulators and they have different features, and takes different configurable parameters. A short list of the current network on Chip simulators includes Noxim {{cite:2a6f176db9ed164b2c98c08514b1d4b1c19031a0}}, NIRGAM {{cite:bdd76af8e82cc4b71f099ca3b7b4eb6cc026d5b5}} and Nostrum NSSE {{cite:96a175f6af0494df6e98ad2ed70c4d99a01a85d1}}. These simulator are cycle accurate NOC simulator developed using SystemC. In Noxim {{cite:2a6f176db9ed164b2c98c08514b1d4b1c19031a0}}, user can customize many parameters likes network size, buffer size, packet size distribution, routing algorithm, packet injection rate, traffic time distribution,traffic pattern, and etc to simulate NoC. NIRGAM {{cite:bdd76af8e82cc4b71f099ca3b7b4eb6cc026d5b5}} is a simulator similar to Noxim for NoC routing and application modeling. Both Noxim and NIRGAM uses randomly generated traffic as input. NNSE takes topology, routing policy, and traffic patterns and based on these configuration parameters a simulator is built and executed. NNSE uses both randomly generated traffic and core generated trace traffic. These simulators mostly design to simulate network on chip target to mainly buffered network and without supporting memory hierarchy. Jantsch et. al. {{cite:f95e95d6a702ff9d03e938cb36209dfd3a8a625b}} explain performance different memory architecture and management in an NoC platform. As most of the available multicore simulators like Multi2sim {{cite:4ce1e4b7ece28c101f7bd001b0560e2206438284}}, SESC {{cite:c4c37400d8e0cc9f3700923d8bd81d12423d61d0}}, SimicGems {{cite:ef8abdc09aa887ee0f37e0ade565612d3b7fa558}}, and, etc are very slow. And NoC simulators written in SystemC are extremely slow. So we can't simulate a bigger design with more than 100 cores. Many researchers propose to enhance NoC simulation using multi-threaded parallel core or accelerated using FPGAs and these are HAsim, Protoflex and FAST {{cite:065209e33527964e54ec2e37513d8f1fd00c0c4b}}, {{cite:2864283847ef9ef2739fd0456fe81813eda44fb6}}, {{cite:85b8a878e8355dbbe20549ee08ae7b9c6132a8c0}}. In SimFlex {{cite:810992a431b4e844cdd8c6061ea363c4efeec491}}, another group of researcher uses statistical sampling of system to speed up the simulation of targeted multicore system. DARSIM {{cite:077b8fbafd3a22d33c7ef6f0706b00bd44b11ab6}}, HORNET {{cite:f32ac3bdae93715973dd0677b548deae9dddde31}} and Graphite {{cite:a3730f7917f1d8803d401570a5c20e75e0dce848}} are cycle accurate parallel simulator targeted to explore multicore architecture. These simulator can be executed either in network only mode by using synthetic traces or in full system multicore mode using a build in MIPS core simulator. These simulator target to run simulation on multicore (power full Xeon like octa or hexa core) server system. Also both the simulator uses several techniques to speed up the simulation. These techniques are direct execution, seamless multicore and multi-machine distribution, and lax or periodic synchronization. Others parallel simulators that target multicore architecture are COTSon {{cite:5ed253859cef2c6d1c4ed433d9bf321b0e4377f6}}, BigSim {{cite:3cb1ce177842105d041dfcf512db26dfa67a64f1}}, FastMP {{cite:541f8cc00b53fa13f8031db623a3324074b5d750}} and Wisconsin Wind Tunnel (WWT) {{cite:538598bcd0581f288564e139ac29e454dbe4e9a0}}.",
"In {{cite:7a3a5f0262c379a4340affcf10efb61a4be180a9}} authors presented first GPU-based NoC simulator without considering bufferless, that is the fastest NoC simulators built till date. To the best of our knowledge, our simulator is the first attempt on building GPU based simulator for LSPD cache system large chip multicore connected with bufferless NoC."
] |
[] |
Modeling emergence of norms in multi-agent systems by applying tipping
points ideas
|
Norms are known to be a major factor determining humans behavior. It's alsoshown that norms can be quite effective tool for building agent-basedsocieties. Various normative architectures have been proposed for designingnormative multi-agent systems (NorMAS). Due to human nature of the conceptnorms, many of these architectures are built based on theories in socialsciences. Tipping point theory, as is briefly discussed in this paper, seems tohave a great potential to be used for designing normative architectures. Thistheory deals with the factors that affect social epidemics that arise in humansocieties. In this paper, we try to apply the main concepts of this theory toagent-based normative architectures. We show several ways to implement theseconcepts, and study their effects in an agent-based normative scenario.
|
[
"Hollander and Wu {{cite:9cd0520b3c4f6b1e7ac54688f4ac20781d032e90}} refer to three categories of normative studies in the social sciences: 1) social function of norms {{cite:3c4e4e3e2beeb4c98759600d718f36451e8ac0d6}}, 2) impact of social norms {{cite:2bf7841ae6e6d3c50724153b2ff40dcbfba75994}}, 3) mechanisms leading to the emergence and creation of norms {{cite:afed04176721f49344d5e42ee3e5d275cf1f4e76}}. In the context of social function, norms are often concerned with the oughtness and expectations of agent behavior; where oughtness refers to the condition where an agent should or should not perform an action, and expectation refers to the behavior that other agents expect to observe from that agent {{cite:0665719cfc3af22020eb6b2ce3662814d6ee5dcc}}, {{cite:8bb740786fb070e860fa1483af668fb6ac7d3538}}. An example of work belonging to this category is Boella and Torre's architecture containing separate subsystems for counts-as conditionals, conditional obligations, and conditional permissions {{cite:e99853e195ae495a24f80232084c7aba22018d84}}.",
"Within the second category, social impact, norms are considered in terms of cost provided to or imposed on the parties involved in a social interaction {{cite:e5ea72159de1f121861603c67d7214ec1974b1b1}}. For instance, punishment and sanctions are introduced as two enforcement mechanisms used to achieve the necessary social control required to impose social norms {{cite:d69dbea5b60d381523ad6503c2a1ddd13ac15ef4}}. Here they demonstrate a normative agent that can punish and sanction defectors and also dynamically choose the right amount of punishment and sanction to impose {{cite:2e2afb2e98ac5f20427afbddfe32e7ee591b7d82}}.",
"As noted in {{cite:9cd0520b3c4f6b1e7ac54688f4ac20781d032e90}}, the third category is concerned with the how of norms more than the why. Relevant work in NorMAS domain that falls into this category is divided into the following groups {{cite:10eff03feb52c00d0d85b565a8705ea18e7715e2}}. The first group is composed of research in which norms are practically hard-wired into the system or dictated directly to the agents. In the second set, norms emerge from social interactions among agents. Sen and Airiau's work {{cite:610169e719a34fe5f6d3e8ce1beb3cf4a803ec70}} in which agent interactions are modeled using payoff matrices, focuses on norm emergence through social learning in agent societies. Conte et al.'s EMIL was designed as an architecture for modeling norm emergence {{cite:7596093074a599cd79d784c711fbe41de0085cf9}}. The EMIL architecture includes a dynamic cognitive model of norm emergence and innovation {{cite:eeda399bca61d2d935d3f720cc5e7f8c29255e56}}.",
"Much of the existing work in normative multi-agent systems explicitly or implicitly relies on social science theories {{cite:e88c792cf2984c8f1cccf12317228ecd12bb5073}}. In a recent work, some of the well-known theories of philosopher David Hume were evaluated using an agent-based model called HUME{{formula:b55d47e4-76a1-469e-988d-2fd753959056}} {{cite:7596093074a599cd79d784c711fbe41de0085cf9}}. This work demonstrates how social justice concepts can even emerge from heterogeneous agents that are not endowed with norm representations.",
"Self-determination theory is also referenced by some of the normative works {{cite:7e0b432bc320584b8e43b570584ae2c7340131a0}}. Here the focus is on the agents' motivation and the extent to which the motivation is intrinsic or extrinsic. Neumann studied existing normative architectures to see how much they comply with self-determination theory {{cite:b1624cfd2d92ec315d30b72c41874144db138cd8}}.",
"Practice theory is an example drawn from anthropology; this theory describes how changes in the society are based on the interactions between the human agents and social structure. For instance, an agent-based model for energy demand and supply social practices is presented in {{cite:06f30ab32b8813020f200c3e567196778f247e09}}, which shows how energy consumption norms form and evolve in urban societies. The next section provides background on tipping point theory, primarily from a sociology perspective."
] |
[] |
Tracking Federated Queries in the Linked Data
|
Federated query engines allow data consumers to execute queries over thefederation of Linked Data (LD). However, as federated queries are decomposedinto potentially thousands of subqueries distributed among SPARQL endpoints,data providers do not know federated queries, they only know subqueries theyprocess. Consequently, unlike warehousing approaches, LD data providers have noaccess to secondary data. In this paper, we propose FETA (FEderated queryTrAcking), a query tracking algorithm that infers Basic Graph Patterns (BGPs)processed by a federation from a shared log maintained by data providers.Concurrent execution of thousand subqueries generated by multiple federatedquery engines makes the query tracking process challenging and uncertain.Experiments with Anapsid show that FETA is able to extract BGPs which, even ina worst case scenario, contain BGPs of original queries.
|
[
"Federated query tracking is related to webtracking {{cite:56cbc0a934d8f8d502f1b7eebe4d5ae8232acafa}}. In web tracking, a first-partywebsite authorizes a third party to learn about its users. Analogously, FETA plays the role of the third party.However, logs collected in federated query tracking is the result ofthe execution of a physical plan in distributed query processingcompared to a more simple web navigation flow in web tracking.",
"Extracting information from raw logs is traditionally a data miningprocess {{cite:86b578a027e1fcbbd380c8bd31bb0c055f5935e0}} involving the following steps:[(i)]",
"data selection identifies thetarget dataset and relevant attributes that will be used to derivenew information;",
"data cleaning removes noise andoutliers, transforms field values to common units, generates newfields and finally brings the data into the structured data schemathat is used for storage, e.g., relational databases, XML;",
"data mining applies data analysis and discovery algorithmsbased on machine learning, pattern recognition, statistics and othermethods;",
"finally evaluation presents the new knowledgein a form that will be also understandable from the end user, e.g.,through visualization.FETA can be located at the data mining step because it transforms raw logsinto a sequence of sets of BGPs. In experiments, we presented oneextraction for a period of time. Repeating extractions will generate asequence of extracted BGPs than can be used for association rulesmining of frequent pattern detection.Conclusions and future workFederated query tracking allows data providers to access secondary data in Linked Data federation. We proposed FETA, a federatedquery tracking approach that extracts original federated Basic Graph Patternsfrom a shared log maintained by data providers. FETA links andunlinks variables present in different subqueries of the federated logby applying a set of heuristics we presented in this paper.Even in a worst case scenario with Anapsid, FETA extracts BGPs that containoriginal BGPs of federated join queries. Extracted BGPs with endpoints'information give valuable information to data providers about which triples are joined, when and by whom.We think FETA opens several interesting perspectives. First,heuristics can be improved in many ways by better using semantics ofpredicates and answers. Second, we conducted experiments on one slice of federatedlog. Repeating BGPs extractions on successive slices allows to applytraditional data mining techniques such as extracting frequentspatterns. Third, we limited experiments to Anapsid. Extending to FedXquery engine {{cite:4a874fab974b04be219ff51a60d0062b3b85aeee}} will challenge proposedheuristics because FedX physical operators produce slightlydifferent query traces."
] |
[] |
Neural Machine Translation of Rare Words with Subword Units
|
Neural machine translation (NMT) models typically operate with a fixedvocabulary, but translation is an open-vocabulary problem. Previous workaddresses the translation of out-of-vocabulary words by backing off to adictionary. In this paper, we introduce a simpler and more effective approach,making the NMT model capable of open-vocabulary translation by encoding rareand unknown words as sequences of subword units. This is based on the intuitionthat various word classes are translatable via smaller units than words, forinstance names (via character copying or transliteration), compounds (viacompositional translation), and cognates and loanwords (via phonological andmorphological transformations). We discuss the suitability of different wordsegmentation techniques, including simple character n-gram models and asegmentation based on the byte pair encoding compression algorithm, andempirically show that subword models improve over a back-off dictionarybaseline for the WMT 15 translation tasks English-German and English-Russian by1.1 and 1.3 BLEU, respectively.
|
[
"For Statistical Machine Translation (SMT), the translation of unknown words has been the subject of intensive research.",
"A large proportion of unknown words are names, which can just be copied into the target text if both languages share an alphabet.If alphabets differ, transliteration is required {{cite:bab83bdbfcb93c20ce35d7168b4eebc48e006120}}.Character-based translation has also been investigated with phrase-based models, which proved especially successful for closely related languages {{cite:d15dfc21197ddb6ab7fb0419471a9f654be01745}}, {{cite:eebfa37e39cf1c098bdae4af6102d8d7e745036a}}, {{cite:f7ab4f4a36b9124576fab4690be9ab52f296e284}}.",
"The segmentation of morphologically complex words such as compounds is widely used for SMT, and various algorithms for morpheme segmentation have been investigated {{cite:4b33e970f93b6784a314b5cef89188d07d03b93f}}, {{cite:febba30a22f7936ff2b2657c2430b40811f8b158}}, {{cite:edcbc9d3c911b533ee5f88ce35073c755dec43a5}}, {{cite:8a1594b7af6c86c9353944f9d2257adc40c6bf4a}}.Segmentation algorithms commonly used for phrase-based SMT tend to be conservative in their splitting decisions,whereas we aim for an aggressive segmentation that allows for open-vocabulary translation with a compact network vocabulary, and without having to resort to back-off dictionaries.",
"The best choice of subword units may be task-specific.For speech recognition, phone-level language models have been used {{cite:cf3d57b426ca799f8f9b911b50628b0b435cc206}}.mikolov2012 investigate subword language models, and propose to use syllables.For multilingual segmentation tasks, multilingual algorithms have been proposed {{cite:84de902c2a7474ebff4e85e4563fcf5e2c6a0671}}.We find these intriguing, but inapplicable at test time.",
"Various techniques have been proposed to produce fixed-length continuous word vectors based on characters or morphemes {{cite:175fdf1ac68e9e5b34ecf3cfba40f93ce920f32f}}, {{cite:473ec89d23792f90867c6aaf4f492da1a047ae94}}, {{cite:0bbaca857fb00c638431115061d5225fc0b72092}}, {{cite:0855c253bc285515042a53dbdbd462b7ebd707d0}}.An effort to apply such techniques to NMT, parallel to ours, has found no significant improvement over word-based approaches {{cite:f30bda3eb407498684e2fd60f1106ca68104eeed}}.One technical difference from our work is that the attention mechanism still operates on the level of words in the model by 2015arXiv151104586L, and that the representation of each word is fixed-length.We expect that the attention mechanism benefits from our variable-length representation: the network can learn to place attention on different subword units at each step.Recall our introductory example Abwasserbehandlungsanlange, for which a subword segmentation avoids the information bottleneck of a fixed-length representation.",
"Neural machine translation differs from phrase-based methods in that there are strong incentives to minimize the vocabulary size of neural models to increase time and space efficiency, and to allow for translation without back-off models.At the same time, we also want a compact representation of the text itself, since an increase in text length reduces efficiency and increases the distances over which neural models need to pass information.",
"A simple method to manipulate the trade-off between vocabulary size and text size is to use shortlists of unsegmented words, using subword units only for rare words.As an alternative, we propose a segmentation algorithm based on byte pair encoding (BPE), which lets us learn a vocabulary that provides a good compression rate of the text."
] |
[] |
A breakpoint detection error function for segmentation model selection
and evaluation
|
We consider the multiple breakpoint detection problem, which is concernedwith detecting the locations of several distinct changes in a one-dimensionalnoisy data series. We propose the breakpointError, a function that can be usedto evaluate estimated breakpoint locations, given the known locations of truebreakpoints. We discuss an application of the breakpointError for findingoptimal penalties for breakpoint detection in simulated data. Finally, we showhow to relax the breakpointError to obtain an annotation error function whichcan be used more readily in practice on real data. A fast C implementation ofan algorithm that computes the breakpointError is available in an R package onR-Forge.
|
[
"This paper has been revised and expanded from Chapter 4 of thedoctoral thesis of {{cite:f24e08b6ba30f5d78ffa9f57470f4c88445ab1c4}}, which has not beenpreviously published elsewhere. Differences include minor changesin notation, an expanded introduction, and more complete references.",
"The main subject of this paper is the breakpointError (defined inSection ), which is a function for preciselymeasuring the breakpoint detection accuracy of a segmentationmodel. There are several other approaches for evaluating segmentationmodels. {{cite:19feb4e4669c80ba4f13e13dac8cd65e83c6da80}} compared the number of detected breakpointswith the number of true breakpoints, ignoring the positions of thebreakpoints. A more precise method was proposed by{{cite:500d590c23a0ed3ee915f6510501ef025138c610}}, who checked if the detected breakpointsappear in regions of arbitrary size around the true breakpoints. Incontrast, the breakpointError we propose in this paper has noarbitrary region size parameter. {{cite:ed7af6ede07834bf038750e2ec63f3e3e668d636}} used exactequality of the estimated and true breakpoint location in theirasymptotic theoretical analysis. The breakpointError function is moreprecise since it is able to quantify that a guess close to a truebreakpoint is better than a guess far from a true breakpoint. A finalclass of methods uses an annotated region database to quantify falsepositive and false negative breakpoint detections{{cite:8a13e60e7eb43d6db0537455acfe99f2c845b842}}, {{cite:d63e697edca6f11bdfaf0166e493294d760ae8e0}}. An annotation databasecan be created by drawing regions on scatterplots of the data using agraphical user interface {{cite:620d78c6ac774bd93ac3a339ac5caa7346f40432}}. Evaluating a segmentationmodel via annotated regions is similar to the breakpointError functionwe propose in this paper, and the precise link between these methodswill be explored in Section .",
"Section shows one example application of thebreakpointError function, for determining the optimal form of penaltyfunctions in segmentation models for simulated data. Many relatedpenalties have been proposed for the change-point detectionproblem. The standard AIC or BIC criteria are not well adapted in thiscontext since the model collection is exponential{{cite:48cd18928c1426cdf6ebb4a950f1bad32082ce3c}}, {{cite:227a35916d98415bb87c4461a8371431b812bea4}}, {{cite:2fcfd462db390526bdf9fde2198571ab5581b416}}, {{cite:b6c7193ffa5bb9b4c0d913a5d5da151e80721062}}, and also because change-points arediscrete parameters {{cite:9f6ca62f05f753a06573b8d96605ce52a820f320}}. Many criteria specifically adaptedto change-point models have been proposed. For example, there aremany different variants of the BIC {{cite:9950303590515ef6d419e66c27a15079d58e40c1}}, {{cite:bea9d84db201ed85585825d7a4af708e0aecd6e2}}, {{cite:9f6ca62f05f753a06573b8d96605ce52a820f320}}, and themodel selection theory of Birgé and Massart suggest other penalties{{cite:54e91a4199a8282d812e93db36cde86cc4736528}}, {{cite:d07deb41aa0cace243af9f24559a317d0af27f73}}, {{cite:48cd18928c1426cdf6ebb4a950f1bad32082ce3c}}, {{cite:bba4c3b4c983dc8ed51bdb3a9d6099ed5532cfd4}}. The precisedifferences between these penalties and the penalties that we findwill be discussed in Section , but the maindifference is that the penalties discussed in this paper arespecifically designed to minimize the breakpointError (rather thansome other function, e.g. the squared error or negative log likelihoodof the data)."
] |
[] |
Effective Spatial Data Partitioning for Scalable Query Processing
|
Recently, MapReduce based spatial query systems have emerged as a costeffective and scalable solution to large scale spatial data processing andanalytics. MapReduce based systems achieve massive scalability by partitioningthe data and running query tasks on those partitions in parallel. Therefore,effective data partitioning is critical for task parallelization, loadbalancing, and directly affects system performance. However, several pitfallsof spatial data partitioning make this task particularly challenging. First,data skew is very common in spatial applications. To achieve best queryperformance, data skew need to be reduced. Second, spatial partitioningapproaches generate boundary objects that cross multiple partitions, and addextra query processing overhead. Consequently, boundary objects need to beminimized. Third, the high computational complexity of spatial partitioningalgorithms combined with massive amounts of data require an efficient approachfor partitioning to achieve overall fast query response. In this paper, weprovide a systematic evaluation of multiple spatial partitioning methods with aset of different partitioning strategies, and study their implications on theperformance of MapReduce based spatial queries. We also study sampling basedpartitioning methods and their impact on queries, and propose several MapReducebased high performance spatial partitioning methods. The main objective of ourwork is to provide a comprehensive guidance for optimal spatial datapartitioning to support scalable and fast spatial data processing in massivelyparallel data processing frameworks such as MapReduce. The algorithms developedin this work are open source and can be easily integrated into different highperformance spatial data processing systems.
|
[
"To the best of of our knowledge, this is the first work that studies the spatial datapartition problem in detail. Data partition problem is discussed extensively in thecontext of database systems in the last few decades {{cite:94e6a7d4de4d7bc226e2f1e6070288f3908fb942}}, {{cite:f276b8b901cc23e9f13dc447fcf8829869830de6}}. Fixed grid partition and its variations are used for spatial join processing in {{cite:37b8f2fdf5caea86f3696117e2c45cddc48fdb19}}, {{cite:1775d60530c141d6a96a00175546b0f869cdc140}}. Le et al. {{cite:123a9fa6a6a4259a58ea1858594cf879d5dda9e3}} studied the problem of finding optimal splitters for largeinterval data. More recently, MapReduce based systems emerged as an effective solution to Spatial Big Data challenges {{cite:e00675372a36930963f19a89f23535ce408cf931}}.HadoopGIS {{cite:d17d937c9bb6b491c51e5ae700565f4eac9231e4}} is a spatial data warehousing system that is based on a general spatial query processing framework. The system uses SQL as the query language, and integrated into Hive{{cite:a4d030cdfda0deb219828bc4dfbb87e3da668321}}. SpatialHadoop {{cite:2d68bbd5d65d02239cbc632040a6f1714c107f15}} is an extension ofHadoop for spatial query processing, and it extends Pig {{cite:060947f27701a7a746c794341ed00ccedfa08120}} at the query language layer. Ray et al. {{cite:e933442fb8e2ddba815fb2c6f1506ac085b339d6}} proposed a spatial data analysis infrastructure that uses a combination of cloud environment and relational database systems. Authors also briefly discussed a hybridapproach that uses Hilbert Curve and space partitioning for spatial join processing.",
"Spatial histogram construction is extensively studied in database settings, and it is widely used for approximate query processing. The main goal of spatial histogram construction is to partition the multi-dimensional data into buckets (most often a bucket represents a rectangular region), where data within buckets is uniformly distributed. In that sense, spatial histogram generation is relevant to spatial partitioning, but not the same. In {{cite:7aaffa410b77c21bede3ca598f893cd04682f446}}, authors have showed that computing the non-overlapping rectangular partitioning with near-uniform data distribution within buckets is NP-hard. One of the pioneering works is {{cite:ea2435acc423e43443b7962c5fb489b1d789ba6b}}, in which authors proposed to extend the concept of equi-depth histogram to multidimensional data. An in-memory data structure hTree is designed for storing the histograms. It constructs non-overlapping partitioning of multidimensional space based on object frequencies. However location of objects are not considered for histogram construction, which may result in skewed histograms. MinSkew histogram {{cite:2def41bd3b747b421c32cf46c12f3a92eb547de1}} is proposed to remedy some of the disadvantages of hTree.Specifically, the authors proposed two construction strategies. The first approach has two phases. In the first phase, the algorithm tiles the spatial universe into uniform regular grids and stores the number of intersecting spatial objects for each tile. Then based on the tiling, a recursive binary space partitioning (BSP) is used for histogram construction. Authors experimentally observed that a fixed-size tiling is sensitive to the size of the queries (high grid resolution favors small sized queries and vice versa), and proposed another approach MinSkew-Progressive-Refinement which can utilize multi-resolution tiling.",
"Spatial histogram construction is extensively studied in spatial database settings, and it is widely used for approximate query processing. The main goal of spatial histogram construction is to partition the multi-dimensional data into buckets (most often a bucket represents a rectangular region), where data within buckets is uniformly distributed. GenHist {{cite:1f1e39c9639feba1695abfc29c85059ba3b26eab}} is a recent approach which can identify high density regions for real valued attributes. However, in GenHist bucket rectangles may overlap, and the buckets can be contained in other buckets. It uses a fixed-size grid as the basis of histogram construction. More recently, an approach called STHist {{cite:010e068ba6635dd8fd381576e370a184375afef6}} is proposed to generate density aware histograms. In the basic STHist algorithm, decision about whether the region is dense is made by applying a sliding window over all dimensions, by approximating the frequency distribution by a marginal distribution.In the advanced variant called STForest, the algorithm first computes coarse partitions according to the object skew, and then applies a sliding window algorithm to them.STHist has a time complexity of {{formula:1fece07b-1d20-4ca7-bdbc-08e1dd0e9e24}} for 2-dimensional and {{formula:41e5e4df-7e49-46c1-b0b6-25c9c2572ec2}} for 3-dimensional data.",
"A convenient approach to obtain a spatial histogram is to generate it using a spatial index structure like R-Tree {{cite:f2780f4d418bc839a9e6d01fc5545b68f5cc6cd3}}, R{{formula:c5769b91-38c7-43a4-9683-465d01776c08}} -Tree {{cite:a09edd869b719074747ee119677352848597f12b}}, R{{formula:fa5f94a6-a433-485d-81a0-a46048dfab4b}} Tree {{cite:6e57c0e589ecb8a51c5889da3cc01f42deed6eed}} etc. RK-Hist {{cite:3ea4e742f64e11414808536f6abf01ce157ebb00}} is an example of such approach which is based on R-tree bulk-loading procedure. The data is presorted according the Hilbert space-filling-curve. After the leaf nodes are generated, a histogram can be generated by packing nodes according to the sorting order in equi-sized histogram buckets. However, this may not necessarily generate a good partitioning. Specifically, for approximately uniformly distributed data equi-sized partitioning wastes buckets for regions with a high object density and produces high overlap between buckets. Therefore, the authors proposed a greedy algorithm utilizing a sliding window of pages along the Hilbert order. The algorithm is parametrized with a number of buckets that should be considered for a split. A bucket-split is applied if it leads to an improvement according to the proposed cost function. More recently, a new approach R-V {{cite:343c490b1a2af4f5302de4dc22304a22f6381d13}} is proposed to overcome skewed-data distribution problem."
] |
[] |
Modal Characterisations of Behavioural Pseudometrics
|
For the model of probabilistic labelled transition systems that allow for theco-existence of nondeterminism and probabilities, we present two notions ofbisimulation metrics: one is state-based and the other is distribution-based.We provide a sound and complete modal characterisation for each of them, usingreal-valued modal logics based on the Hennessy-Milner logic. The logic forcharacterising the state-based metric is much simpler than an earlier logic byDesharnais et al. as it uses only two non-expansive operators rather than thegeneral class of non-expansive operators.
|
[
"Metrics for probabilistic transition systems are first suggested byGiacalone et al. to formalizea notion of distance between processes. They are used also in{{cite:48d229fa0ee345ff08a34f77eb77a4ae3ea52fb8}}, {{cite:ef05e6612ac233140f8e08f5d5acfe7923847478}} to give denotational semantics for reactivemodels. De Vink and Rutten {{cite:e37c128991b6312fa451da0726bf2dfe6e162ba3}} show that discreteprobabilistic transition systems can be viewed as coalgebras. Theyconsider the category of complete ultrametric spaces. Similarultrametric spaces are considered by den Hartog in {{cite:7e5496293392884383b8d8abbc49654e21fd9224}}.",
"Metrics for deterministic systems are extensively studied.Desharnais et al. {{cite:ee33256ab41032e8af577edc322d71f5015db5ac}} propose a logical pseudometricfor labelled Markov chains, which is a reactive model of probabilisticsystems. A similar pseudometricis defined by van Breugel and Worrell{{cite:46e71adb78ad70dd8d23ca0250665d4de0106b26}} via the terminal coalgebra of a functor based on a metricon the space of Borel probability measures. The metric of {{cite:ee33256ab41032e8af577edc322d71f5015db5ac}}, {{cite:f6ee891805b0a09345e4d22ecfc1e149f4b86c34}} works forcontinuous probabilistic transition systems, while in this work weconcentrate on discrete systems.Interestingly, vanBreugel and Worrell {{cite:682d8021363b7d050aad194dc527e0b840b1ae05}} also present a polynomial-timealgorithm to approximate their coalgebraic distances. Furthermore, van Breugel et al. propose an algorithm to approximate a behavioural pseudometric without discount {{cite:d5479c3a8dfc4c0f1b15b11b3323017e20217ae9}}. In {{cite:2f991885264fdd3bc6abc9b5d7940afb9b7eadad}} a sampling algorithm for calculating bisimulation distances in Markov decision processes is shown to have good performance.In {{cite:241b896d940afaf9b3a7a43fc9392ba341e5ba69}}, {{cite:080c43d635960f7e35b1b9a5454b2d15c3fec61a}} the probabilistic bisimulation metric on game structures is characterised by a quantitative {{formula:e0d8282d-a1cb-42bf-9efa-71ba65bd931a}} -caluclus. Algorithms for game metrics are proposed in {{cite:5f3c3aeb13a686aa2c83c60d170c6a575c560951}}, {{cite:dc271dd2fe3d63766ba1828c8d5d2a4fae37d206}}.",
"Metrics for nondeterministic probabilistic systems are considered in{{cite:5670cc12426838b178806d3020aed6f983bfa5b7}}, where Desharnais et al. deal with labelled concurrentMarkov chains (similar to pLTSs, this model can be captured by the simple probabilisticautomata of {{cite:0f6ccedaadc93a51b3ba3a410ff3e47e8a44a79b}}). They show that the greatest fixed pointof a monotonic function on pseudometrics corresponds to the weakprobabilistic bisimilarity of {{cite:a9d8bdaa55e50542c37862e075907695ef255319}}.",
"In {{cite:3024c4935d062973af01223925caa4e2fa51ad9e}} a notion of trace metric is proposed for pLTSs and a tool is developed to compute the trace metric. In {{cite:d34bef6a632915386a092c80fe840b8be1f8db4a}} a notion of bisimulation metric is proposed that extends the approach of {{cite:5670cc12426838b178806d3020aed6f983bfa5b7}}, {{cite:ee33256ab41032e8af577edc322d71f5015db5ac}} to a more general framework called action-labelled quantitative transition systems.",
"In {{cite:dc67957f8c5ae76345dd386dcfde8dbc55ba66a3}} de Alfaro et al. consider metric transition systems in which the propositions at each state are interpreted as elements of metric spaces. In that setting, trace equivalence and bisimulation give rise to linear and branching distances that can be characterised by quantitative versions of linear-time temporal logic {{cite:d8acfb1c7ada644909c7e3e578c6e707f58ed484}} and {{formula:9ff754f7-a93b-4844-932b-2d850fb32e5b}} -calculus {{cite:c55979d7004db01652bfab82284343aaedf71003}}.",
"In {{cite:ad993f28be19d14d218f26bd33e088fbf974943d}} Ying proposes anotion of bisimulation index for the usual labelled transition systems,by using ultrametrics on actions instead of using pseudometrics onstates. He applies bisimulation indexes on timed CCS and real timeACP. But the deeper connection between {{cite:ad993f28be19d14d218f26bd33e088fbf974943d}} and our workworths some further studies.",
"In {{cite:812b90d5526f165db87fd4a1033e2f3efcb93359}} a notion of uniform continuity is proposed to be an appropriate property of probabilistic processes for compositional reasoning with respect to bisimulation metric semantics."
] |
[] |
New bounds on curve tangencies and orthogonalities
|
We establish new bounds on the number of tangencies and orthogonalintersections determined by an arrangement of curves. First, given a set of $n$algebraic plane curves, we show that there are $O(n^{3/2})$ points where two ormore curves are tangent. In particular, if no three curves are mutually tangentat a common point, then there are $O(n^{3/2})$ curve-curve tangencies. Second,given a family of algebraic plane curves and a set of $n$ curves from thisfamily, we show that either there are $O(n^{3/2})$ points where two or morecurves are orthogonal, or the family of curves has certain special properties. We obtain these bounds by transforming the arrangement of plane curves intoan arrangement of space curves so that tangency (or orthogonality) of theoriginal plane curves corresponds to intersection of space curves. We thenbound the number of intersections of the corresponding space curves. For thecase of curve-curve tangency, we use a polynomial method technique that isreminiscent of Guth and Katz's proof of the joints theorem. For the case oforthogonal curve intersections, we employ a bound of Guth and the third authorto control the number of two-rich points in space curve arrangements.
|
[
"In {{cite:de282d397ffe48a7b3ace8fbda49670b32786d2d}}, Clarkson et. al. developed techniques to bound the number of incidences between points and surfaces in {{formula:18c5715c-50f9-41b5-b3c8-0bfc7f08bd64}} . In {{cite:97b1d03fa18d17cc13b77828953c6cfc931bf221}}, {{cite:fc8a6612fc830a390f7941c058d122d132cbf59f}}, Wolff observedThe tangency bound for circles is not stated explicitly in {{cite:97b1d03fa18d17cc13b77828953c6cfc931bf221}}, {{cite:fc8a6612fc830a390f7941c058d122d132cbf59f}}, but they are discussed in {{cite:697cbe6bb52e273917e992e3772f638a027f3402}}, which is an expository paper that discusses the results in {{cite:97b1d03fa18d17cc13b77828953c6cfc931bf221}}, {{cite:fc8a6612fc830a390f7941c058d122d132cbf59f}} that these techniques could be used to bound the number of tangencies between circles in {{formula:51b086aa-970c-4f92-ba9e-d86a7a092d44}} . Specifically, {{formula:989d4edf-d1b7-4e1b-8a27-95ef26ba1510}} circles in {{formula:3c7aadb1-7a33-4547-a157-c94e197c12fd}} determine {{formula:52cf43f8-bad3-4670-833b-12e0ce5b5213}} directed points of tangency, where {{formula:022607b8-9776-4862-a76c-8d2da25baf22}} is a very slowly growing function. Some of the methods from {{cite:6c7a83342fd998e963201160763785b91b4b76ca}} can also be used to slightly improve this tangency bound to {{formula:41a82122-ecc2-44bf-b3f7-b8ffd607039a}} . All of these methods, however, made crucial use of two facts. First, circles in the plane have three degrees of freedom—a circle can be described by three parameters. While these techniques can be extended to general degree {{formula:f3988c88-1cd8-4289-9e39-df9712ed5155}} algebraic curves in {{formula:da6fedbc-eb73-4784-bad2-0b9c04aaafc0}} , the resulting exponent in the bound becomes worse as {{formula:db31775d-6513-4103-ae25-95e29bfee701}} increases. Second, these older techniques make crucial use of various “cuttings” or “polynomial partitioning” theorems, and these results generally cannot be extended to fields other than {{formula:51b8e5f4-d502-42a2-ab1a-a7be2e6f8a8c}} . Theorem REF suffers from neither of these constraints.",
"In {{cite:2ab81c1e86e4b4d0edbe83deff1c41de3befb54d}}, Megyesi and Szabó proved that if {{formula:eceaeb6e-d9c3-49b1-a200-0dacda450bcf}} is a field whose characteristic is not two, then {{formula:abeb983f-6b01-4eea-95e7-b59e5ce6057e}} conics in {{formula:6ef4b130-a6e4-4956-b0c2-c4b47fd0c157}} determine {{formula:9ad8242d-20ce-48de-88c4-c7edaa5635ac}} directed points of tangency. Theorem REF improves this to {{formula:86ffbb6d-9a26-4922-bb48-aa2e19dee0f4}} in the special case where {{formula:5023b790-a890-4ed2-ac72-cf3faaa66aa6}} or {{formula:396ad6d3-d629-42ba-8be6-28e21a78e327}} for a suitable (absolute) constant {{formula:d3668cf8-565d-4c13-9b6d-d31e30f2e956}} .",
"In {{cite:a452edc95f649e982eb8af755e3b61b40ee7cdf2}}, Marcus and Tardos showed that any family of {{formula:cfaf34dc-2ad2-47dd-93bd-c39d614f30c1}} pseudocircles determines {{formula:5e84a0c8-d98a-4e19-87e2-e7950b7b4900}} tangencies. In the case where the pseudocircles are defined by bounded-degree algebraic curves, Theorem REF improves this bound. This is also Problem 14 in Chapter 7.1 of Brass-Moser-Pach {{cite:515f965434de81c80f04472a03f9993ea30a6df2}}.",
"Purely combinatorial methods work well for bounding pseudocircle tangencies. However, one can not expect general results for “pseudocubics” as the following simple construction shows. In Figure REF we present an arrangement of curves in {{formula:533706dd-43f6-44e9-8880-a3993ae78ace}} where every two curves meet in two points and every two curves are tangent; this results in {{formula:284b7d38-7bce-47df-b5c9-dd0d60dec7bf}} directed points of tangency. Thus the requirement that the curves be algebraic (and of controlled degree) is crucial.",
"{{formula:8a6429d6-7062-4fe7-95f4-561282bf1a2c}}{{figure:d6aaf159-e128-4146-875c-50a59d6fb279}}"
] |
[] |
Exact Real Arithmetic with Perturbation Analysis and Proof of
Correctness
|
In this article, we consider a simple representation for real numbers andpropose top-down procedures to approximate various algebraic and transcendentaloperations with arbitrary precision. Detailed algorithms and proofs areprovided to guarantee the correctness of the approximations. Moreover, wedevelop and apply a perturbation analysis method to show that our approximationprocedures only recompute expressions when unavoidable. In the last decade, various theories have been developed and implemented torealize real computations with arbitrary precision. Proof of correctness forexisting approaches typically consider basic algebraic operations, whereasdetailed arguments about transcendental operations are not available. Anotherimportant observation is that in each approach some expressions might requireiterative computations to guarantee the desired precision. However, no formalreasoning is provided to prove that such iterative calculations are essentialin the approximation procedures. In our approximations of real functions, weexplicitly relate the precision of the inputs to the guaranteed precision ofthe output, provide full proofs and a precise analysis of the necessity ofiterations.
|
[
"An implementation for a bottom-up approach to exact real arithmetic is proposed in {{cite:d3ab760afd114595389d00e910c5114b3f96b3fd}}. For a given expression, the inputs are computed with predefined precisions and a bottom-up scheme is used to determine the guaranteed precision of the output. Iterative computations are required if the obtained precision is not adequate. A formalization of a top-down approach in a theorem prover is proposed in {{cite:894158d710855a414650d7ea1cc11b86c1bbbff3}}. The author first provides a definition for a metric space based on a ball relation. Afterwards, real numbers are defined as the completion of the metric space {{formula:56c1f149-2d15-4a59-99a5-7b97b975bf31}} . Rational operations are lifted to approximate operations on real numbers. This approach is optimized in {{cite:057dd96434fc3554853b40db089fff42fd115681}}.",
"Two closely-related top-down approaches based on absolute errors have been studied in {{cite:732f1e385e469edb69906cce490588e60eff58c4}}, {{cite:9529249ab2d2f97c523d5e319f250b923cb09750}}. These approaches mainly differ in their approximations of the transcendental functions. In {{cite:732f1e385e469edb69906cce490588e60eff58c4}} the authors introduce a general way for calculating with Taylor expansions and apply this method to approximate the transcendental functions. In {{cite:9529249ab2d2f97c523d5e319f250b923cb09750}}, the approximations of the transcendental functions are treated separately and in a more ad-hoc way. In contrast, our approach is based on relative errors. We provide detailed proofs of correctness for each operation and use perturbation analysis to identify essential recomputations.",
"In {{cite:6c0413b0a0e91d29c8327c1e15283bc3ece8a365}} the authors propose a layered framework for computations with real numbers. The lowest layer is an implementation of floating point arithmetic. In the second layer, arithmetic operations are approximated using polynomial models. The highest layer supports more advanced features such as differential operations. Proof of correctness in Coq and an implementation based on {{cite:6c0413b0a0e91d29c8327c1e15283bc3ece8a365}} are also available. In this article, we have focused on approximating specific operations, whereas in {{cite:6c0413b0a0e91d29c8327c1e15283bc3ece8a365}} polynomial models are discussed in an abstract way without concrete examples from well-known arithmetic operations.",
"The approach introduced in {{cite:904d3d4ad8f1aebf0a5d6909dfa2ee0af7fab67f}} is based on linear fractional transformations (LFTs). Computations are encoded as trees of LFTs; various operations are defined to extract the result of a computation from the corresponding tree. However, this approach does not specify a top-down scheme to relate the desired precision in the output and the required operations on the expression tree. Expression trees are evaluated using lazy evaluation; computations terminate when adequate information is available in the root of a tree.",
"A symbolic approach to exact real arithmetic has been proposed in {{cite:c1fa40885164ce85f6d37e4e2c2c15a3fe44c763}}. The author uses infinite binary sequences in the golden ratio base to represent real numbers. To calculate an expression, first the symbolic techniques available in Maple are applied to obtain a simplified expression. Additional Maple procedures are implemented by the author to extract binary sequences from simplified expressions. Performing operations on binary sequences is also possible. However, choosing a suitable balance between symbolic computations and direct manipulation of binary sequences depends on the given expression. As indicated in {{cite:c1fa40885164ce85f6d37e4e2c2c15a3fe44c763}}, using this approach to its full potentials requires expertise in Maple. Moreover, the procedure might need adaptations for each problem."
] |
[] |
Adaptive-Robust Control of a Class of Uncertain Nonlinear Systems
Utilizing Time-Delayed Input and Position Feedback
|
In this paper, the tracking control problem of a class of Euler-Lagrangesystems subjected to unknown uncertainties is addressed and an adaptive-robustcontrol strategy, christened as Time-Delayed Adaptive Robust Control (TARC) ispresented. The proposed control strategy approximates the unknown dynamicsthrough time-delayed logic, and the switching logic provides robustness againstthe approximation error. The novel adaptation law for the switching gain, incontrast to the conventional adaptive-robust control methodologies, does notrequire either nominal modelling or predefined bounds of the uncertainties.Also, the proposed adaptive law circumvents the overestimation-underestimationproblem of switching gain. The state derivatives in the proposed control law isestimated from past data of the state to alleviate the measurement error whenstate derivatives are not available directly. Moreover, a new stability notionfor time-delayed control is proposed which in turn provides a selectioncriterion for controller gain and sampling interval. Experimental result of theproposed methodology using a nonholonomic wheeled mobile robot (WMR) ispresented and improved tracking accuracy of the proposed control law is notedcompared to time-delayed control and adaptive sliding mode control.
|
[
"It can be observed that TDE degrades tracking performance of both the TDC and F-TDC in the face of uncertainties. The control methods that attempt to counter uncertainties, as reported in {{cite:e8a39d029ff7207c361a8bcec6edbb8a4c0fb8a0}}, {{cite:4538f71bb9c8aaec68f25f3d0c08837af4c180b8}}, {{cite:a45661dc5431bca3a5970c7320d1399d649aaa5f}}, {{cite:9042e45dffb7aa6ca416bd2b3562ed22ff3fd222}}-{{cite:77020220d3df2fd6d815047eb9a164d45ce648d4}}, requires predefined bound on the uncertainties which is not always possible in practical circumstances. To circumvent this situation Adaptive Sliding Mode Control (ASMC) was proposed in {{cite:4bd914a991ef56b999a60be81a653e685af409dd}}-{{cite:9e42edf5b6d54de6575ef3b8a0a7b3f378c7a038}}. The control input of ASMC is given by,{{formula:e0adb2ef-8771-41cf-84af-ffac0fcebab4}} ",
"where, {{formula:f3b362fe-afb3-49ce-8dbd-68c47add72ff}} and {{formula:21873115-afc4-462b-9479-26b9bf7fd6bc}} is the nominal values of {{formula:04450d44-94e5-4a7e-87e5-44ceb62f549b}} and {{formula:144e0e1f-68e3-4a42-99bd-e07a43febaa5}} , and {{formula:da566c49-88b2-43f7-9ea5-bdaca32487ef}} is the switching control input. For a choice of sliding surface {{formula:8b2a5ee9-df85-4d23-a202-2c06a2bb5531}} , {{formula:a2fb699a-950b-45ea-9888-f56d49288886}} and {{formula:bb44ff1f-871d-4945-8647-3cbe8200ede8}} is defined as follows:{{formula:56fd7652-ada2-4d9d-954e-99f1c6e04633}} ",
"The switching control {{formula:694a1f8f-97b4-46ed-a58d-257fa8375279}} is calculated as{{formula:7bdf854a-f128-46ec-b539-c3764e1730b6}} {{formula:7d0e7758-c72e-4951-a2e6-abb6774a0e84}} ",
"where, {{formula:b43e19d2-1084-4617-b301-074ff5b5a038}} is the switching gain, {{formula:464510af-1723-4b56-8ecf-1b2713c9210f}} is a scalar adaptive gain, {{formula:948576e6-672c-4746-a454-8204c73c69fc}} is a threshold value, {{formula:2f9aca85-500c-412f-aaf1-5e37b4e39355}} is small scalar to always keep {{formula:d2e5017a-4d06-4799-b817-9dd388fa33fc}} positive. Evaluation of {{formula:ffcf5102-0944-4fef-b873-243abcc23440}} can be done in two ways {{cite:4bd914a991ef56b999a60be81a653e685af409dd}}:{{formula:c3773930-0e22-4df2-af7e-1653320d44f0}} ",
"where, {{formula:6ec357ae-aab6-49e2-b3fe-1b62703dec46}} is a scalar, {{formula:211cd442-77c7-410b-8534-72d34944f2f2}} is the sampling interval. The choice (REF ) requires predefined bound of uncertainties. It can be noted from (REF ) that even if {{formula:01d494d7-ad84-4265-9378-c464bd7cb5fa}} decreases (resp. increases), unless it falls below (resp. goes above) {{formula:2f5d299a-44f7-405f-bdf8-02c94fa73504}} switching gain does not decrease (resp. increase). This causes overestimation (resp. underestimation) of switching gain and controller accuracy is compromised. Again, improper and low choice of {{formula:85c6a4c6-af2b-4f9b-8508-712fdccc5c9d}} may lead to very high switching gain and consequent chattering. On the other hand, method () assumes that the nominal value of the uncertainties are always greater than the perturbations. This assumption may not hold due to the effect of unmodelled dynamics and thus, necessitates rigorous nominal modelling of the uncertainties in {{formula:cc7d8c6f-fcbf-4643-9391-6e3a8a8e5594}} to design the control law. Either of the two situations, i.e. bound estimation or uncertainty modelling, is not always feasible in practical circumstances and consequently compromises the adaptive nature of the controller."
] |
[] |
Fast Parallel Fixed-Parameter Algorithms via Color Coding
|
Fixed-parameter algorithms have been successfully applied to solve numerousdifficult problems within acceptable time bounds on large inputs. However, mostfixed-parameter algorithms are inherently \emph{sequential} and, thus, make nouse of the parallel hardware present in modern computers. We show that parallelfixed-parameter algorithms do not only exist for numerous parameterizedproblems from the literature -- including vertex cover, packing problems,cluster editing, cutting vertices, finding embeddings, or finding matchings --but that there are parallel algorithms working in \emph{constant} time or atleast in time \emph{depending only on the parameter} (and not on the size ofthe input) for these problems. Phrased in terms of complexity classes, we placenumerous natural parameterized problems in parameterized versions of AC$^0$. Ona more technical level, we show how the \emph{color coding} method can beimplemented in constant time and apply it to embedding problems for graphs ofbounded tree-width or tree-depth and to model checking first-order formulas ingraphs of bounded degree.
|
[
"There is a growing body of literature reporting on the practicalitiesof implementing fixed-parameter algorithms inparallel {{cite:9277bb50748a60c3132a50fb75192a4b0e29bb8f}}.In contrast, there are only fewresults addressing parallel fixed-parameter tractability on atheoretical level (as we do in the present paper), see for instanceCesati and Di Ianni {{cite:898b5e0f5bd8cef5d51370a8fdb790eb9edac741}}.Since it is well-known fromclassical complexity theory that problems solvable in logarithmicspace can be parallelized well, previous research onparameterized logarithmic space contributes to ourunderstanding of which parameterized problems can be parallelized inprinciple. This research was started by Cai, Chen, Downey, andFellows {{cite:f42b46b0abe5df1b602665d206dfb9b883e1df27}}. First (quite technical) complete problemsfor parameterized logarithmic space where later introduced by Chen,Flum, and Grohe {{cite:f77993c48df0a2709a18da7483f6fadd922a2722}}, and by Flum andGrohe {{cite:a118ddf49488ea3d6b8904ec75404876ed4ae956}}. A more structural study of parameterized spaceand circuit classes (which addresses parallelization more directly)was later made by Elberfeld and the last twoauthors {{cite:1aac98dc4653dc9367b4d9c30dbb17773293b81b}}.Parameterized Circuit Complexity wasalso studied by Downey et al. with respect to the WeftHierarchy {{cite:c76aeba8e4deff72d8d4a0c3ca65330e1d884d96}}.Recently,Chen and Müller {{cite:290b838653df2d89331a0483dd63b29411162524}} connected color coding andparameterized space in an algorithm for finding embedding of boundedtree-depth graphs in parameterized logarithmic space (a result whichwe strengthen considerably in Corollary REF ).",
"The first use of the color coding technique can be traced back toAlon, Yuster, and Zwick {{cite:5dc60d0ee75de43d42b8529f4a8785f10f64e701}}. They used the technique toprovide an {{formula:c5edfed0-0455-44e3-8861-fcd750137616}} -algorithm that decides whether there is anembedding of a graph {{formula:d1b21c7d-1a3d-40b6-8c0e-4134299df1a1}} of bounded tree-width into anothergraph {{formula:de190686-0d9c-4766-917f-6a12ad43e36a}} , where {{formula:62128356-87ae-4a91-8de5-6a70c0b03050}} is the parameter."
] |
[] |
Sparsity-based Correction of Exponential Artifacts
|
This paper describes an exponential transient excision algorithm (ETEA). Inbiomedical time series analysis, e.g., in vivo neural recording andelectrocorticography (ECoG), some measurement artifacts take the form ofpiecewise exponential transients. The proposed method is formulated as anunconstrained convex optimization problem, regularized by smoothed l1-normpenalty function, which can be solved by majorization-minimization (MM) method.With a slight modification of the regularizer, ETEA can also suppress moreirregular piecewise smooth artifacts, especially, ocular artifacts (OA) inelectroencephalog- raphy (EEG) data. Examples of synthetic signal, EEG data,and ECoG data are presented to illustrate the proposed algorithms.
|
[
"Some recent works recover signals with transients by various algorithms.In {{cite:a3f38a3cec0882728456ca72bd6bc56fc634c6f9}}, a slowly varying signal is modeled as a local polynomial and an optimization problem using Tikhonov regularization is formulated to capture it.In {{cite:2a0e175a9da2b29a152a800b1ad05cca01762048}}, the slowly varying trend is modeled as a higher-order sparse-derivative signal (e.g., the third-order derivative is sparse).",
"Instead of estimating the slowly varying component via regularization, the {{formula:4a11738c-8868-478b-b0e9-8d6e25b6e0e5}} method {{cite:67d95ffd84b6bfc069882107efe09f6bed5d0553}}estimates a lowpass component by LTI filtering and a piecewise constant component by optimization.In this case, an optimization problem is formulated to estimate the piecewise constant component.The approach proposed here uses a similar technique to recover the lowpass component,but in contrast to {{formula:1cb57793-21a0-496a-8f60-3f929de5bce5}} , it is more general— the regularization is more flexible with a tunable parameter, so that {{formula:63004be2-9f41-4ee0-b71d-8a258b06b629}} can be considered as a special case.",
"Another algorithm related to the approach taken in this paper is the transient artifact reduction algorithm (TARA) {{cite:f0e5fe6131ae3f0baa2b377936bbeb47e02a2b48}}which is utilized to suppress additive piecewise constant artifacts and spikes (similar to a hybrid of Type 2 and Type 3 artifact).The approaches proposed in this work target different types of artifacts (Type 0 and Type 1)and applied in different applications.",
"The irregularity of Type 0 transients leads to a more complicated artifact removal problem, where the artifact are irregular fluctuations.A typical example in EEG is ocular artifacts (OA) caused by the blink and/or movement of eyes.To suppress OA, there are approaches based on empirical mode decomposition (EMD){{cite:302c21cc17c19ae3d171e8ac988930b23b57342e}}, {{cite:c915288c2ef91ca2ee3fd12abd37480d68ffebcd}}, {{cite:6377c56f9a2e45764f611a9fbe830e230dfeba9e}}, {{cite:1bd0738e5bf8f190dc97e36305bc68864d9d0155}},and on independent component analysis (ICA) methods{{cite:b3e4c26cc07c26d81c36c0a8bab37585a575a28d}}, {{cite:bdc7519cc1fb0abb576158b44487d83030d3e49f}}, {{cite:0fad3d397647096525698a191a207d3b133b83a7}}, {{cite:17475a405d0055adee6fc10411e85948d73e071a}}, {{cite:9dd5600248f24d6443b3c6772c853c37595ee922}}, {{cite:1855634454c4ad8bab35c0ea57d131f45a20903e}}.The concept of spatial-frequency in acoustic analysis is also used to remove OA from multichannel signals{{cite:bfaceeaa462efed6d459fbf70b41f6faad3526c4}}, {{cite:97ea5c79bb5a44474965f8e5fd5f6f593a525664}}.In this work, we present a new method to suppress ocular artifacts by proposing a specific model and using sparse optimization.",
"This paper adopts a regularizer inspired by the generalized 1-D total variation {{cite:ec5a50bd1af2840e9e2f44207f0da2da6a5ff107}},wherein the derivative operator in conventional total variation regularizer is generalized to a recursive filter.The regularizers adopted in {{formula:da1b8a9a-51be-45d9-87bb-ad6645e4063e}} and second-order {{formula:662879c2-6377-4303-81c8-9a4d17f6be68}} coincide with first-order and second-order cases of generalized 1-D total variation, respectively.Some differences to the problem discussed in {{cite:ec5a50bd1af2840e9e2f44207f0da2da6a5ff107}} are as follow.Firstly, the signal model (REF ) allows a lowpass baseline as a component,hence, {{formula:8be19ca6-ab5c-4129-ab57-6c2a6ad37f4f}} can be seen as a combination of conventional LTI filtering and generalized 1-D total variation.Secondly, we consider a formulation in terms of banded matrices, for computational efficiency.Thirdly, we give optimality conditions of the proposed problems,and use these conditions as a guide to set the regularization parameters."
] |
[] |
PRAGMA-ENT: Exposing SDN Concepts to Domain Scientists in the Pacific
Rim
|
The Pacific Rim Application and Grid Middleware Assembly (PRAGMA) is aninternational community of researchers that actively collaborate to addressproblems and challenges of common interest in eScience. The PRAGMA ExperimentalNetwork Testbed (PRAGMA-ENT) was established with the goal of constructing aninternational software-defined network (SDN) testbed to offer the necessarynetworking support to the PRAGMA cyberinfrastructure. PRAGMA-ENT is isolated,and PRAGMA researchers have complete freedom to access network resources todevelop, experiment, and evaluate new ideas without the concerns of interferingwith production networks. In the first phase, PRAGMA-ENT focused on establishing an international L2backbone. With support from the Florida Lambda Rail (FLR), Internet2,PacificWave, JGN-X, and TWAREN, PRAGMA-ENT backbone connects Open\-Flow-enabledswitches at University of Florida (UF), University of California San Diego(UCSD), Nara Institute of Science and Technology (NAIST, Japan), OsakaUniversity (Japan), National Institute of Advanced Industrial Science andTechnology (AIST, Japan), and National Center for High-Performance Computing(Taiwan). The second phase of PRAGMA-ENT consisted of evaluation of technologies forthe control plane that enables multiple experiments (i.e., OpenFlowcontrollers) to co-exist. Preliminary experiments with FlowVisor revealed somelimitations leading to the development of a new approach, called AutoVFlow.This paper will share our experience in the establishment of PRAGMA-ENTbackbone (with international L2 links), its current status, and control planeplans. Discussion on preliminary application ideas, including optimization ofrouting control; multipath routing control; and remote visualization will alsobe discussed.
|
[
"SDN concepts are a newly established paradigm that features the separation of the control plane from the data plane {{cite:49a6632d2a08bd2caf638fbc73a973a7a457f20e}}. The control plane has various abstractions that enable administrators/users to control how information flows through the data plane {{cite:8e020e9c811ccc2ccbecfc4e86d6a8da66aedf3f}} by exposing a programmable interface. One advantage of this is that the controllers have a global view of the infrastructure topology and network state, thus providing unprecedented control over the network, which increases the efficiency, extensibility, and security of the network. OpenFlow is an open standard protocol that enables the control plane to interact with the data plane {{cite:36653dd450fda120d4d80e7ed16dac924c3426a9}}.",
"Advanced Layer 2 Service (AL2S) offered by Internet2 is the most successful network service using SDN technologies {{cite:7c75af6e7ff7ccca70b0f65142f422bb393bc620}}, {{cite:bef8d8374928ffb52e5c9676c9b32112205ae74b}}. AL2S provides a dynamic Layer 2 provisioning service across the Internet2 backbone network. Users of Internet2 can provision point-to-point VLANs to other Internet2 users using a Web portal for AL2S. The software talks to an OpenFlow controller to push OpenFlow rules into the networking devices and allocates a VLAN between users across the network backbone in a very dynamic manner. Since the OpenFlow rules perform any necessary VLAN translation on any of the networking nodes in the path, it eliminates the long provisioning delay for matching VLAN IDs over the WAN. This service is very useful to provision a dedicated network for collaborative research, but the AL2S itself does not provide an opportunity for the users to deploy their own OpenFlow controller. Thus, the users do not access and control each networking switch directly.",
"Research Infrastructure for large-Scale network Experiments (RISE) offered by JGN-X is a very unique network service that allows its users to deploy their own OpenFlow controller on the network testbed {{cite:2e5f60966ad0c10afbcc658f877516e16f35bdaa}}; therefore, users can fully control the assigned testbed using their controller. RISE creates multiple virtual switch instances (VSI) on each switch device and assigns them for users. Each VSI acts as a logical OpenFlow switch and is assigned to the user controller. This design is ideal for the PRAGMA testbed since we can have our own OpenFlow controller for the testbed. Internet2 is also beginning a similar service, called network virtualization service (NVS) that allows users to bring their own controller for the testbed, but it is not open for users at this time. We therefore decided to use RISE services as our backbone, and develop our data plane by connecting our OpenFlow switches deployed at participating sites to the backbone."
] |
[] |
How to Verify a Quantum Computation
|
We give a new theoretical solution to a leading-edge experimental challenge,namely to the verification of quantum computations in the regime of highcomputational complexity. Our results are given in the language of quantuminteractive proof systems. Specifically, we show that any language in$\mathsf{BQP}$ has a quantum interactive proof system with a polynomial-timeclassical verifier (who can also prepare random single-qubit pure states), anda quantum polynomial-time prover. Here, soundness is unconditional--i.e., itholds even for computationally unbounded provers. Compared to prior workachieving similar results, our technique does not require the encoding of theinput or of the computation; instead, we rely on encryption of the input(together with a method to perform computations on encrypted inputs), and showthat the random choice between three types of input (defining a computationalrun, versus two types of test runs) suffices. Because the overhead is very lowfor each run (it is linear in the size of the circuit), this shows thatverification could be achieved at minimal cost compared to performing thecomputation. As a proof technique, we use a reduction to an entanglement-basedprotocol; to the best of our knowledge, this is the first time this techniquehas been used in the context of verification of quantum computations, and itenables a relatively straightforward analysis.
|
[
"The complexity class {{formula:f63d71a3-8ea9-4f59-8c7f-1b9ea47b3bb7}} , corresponding to quantum prover interactive proof systems, was originally defined by Aharonov, Ben-Or and Eban {{cite:770d52cf7db37edfa831340270af1783aa0a5fa3}}, who, using techniques from {{cite:b828601935cee1dea0d97c0cdf294fcf4a243fbd}}, showed that {{formula:1bd21f53-1e02-472f-84c6-28251c0ce5a5}} for a verifier with the capacity to perform quantum computations on a constant-sized quantum register (together with polynomial-time classical computation). The main idea of {{cite:770d52cf7db37edfa831340270af1783aa0a5fa3}} is to encode the input into a quantum authentication code {{cite:bb9a1f87855a019b59b490dc5825544d866aa482}}, and to use interactive techniques for quantum computing on authenticated data in order to enable verification of a quantum computation.This result was revisited in light of foundations of physics in {{cite:ad181ba4bae188424e24a17df495bfc71dd62fc1}}, and the protocol was also shown secure in a scenario of composition {{cite:e840756d50c5b53f74ab3cc2b96c8cc14131bed2}}.",
"In a different line of research, Kashefi and Fitzsimons {{cite:1eb91c53a6bbc1e4540227d9d400388fe4a2f4c7}} consider a measurement-based approach to the problem, giving a scheme that requires the verifier to prepare only random single qubits: the main idea is to encode the computation into a larger one which includes a verification mechanism, and to execute the resulting computation using blind quantum computing {{cite:17837d66b991ff739fb81d2dc1700bc9750f4311}}. Thus, success of the encoded computation can be used to deduce the correctness of actual computation. A small-scale version of this protocol was implemented in quantum optics {{cite:ba60c106e4ae57cd677e3d7740fc9381a0a85397}}. Further work by Kapourniotis, Dunjko and Kashefi {{cite:f0deeefb4a23394ba1c68cedd5d7f5fc90425e89}} shows how to combine the {{cite:770d52cf7db37edfa831340270af1783aa0a5fa3}} and {{cite:1eb91c53a6bbc1e4540227d9d400388fe4a2f4c7}} protocols in order to reduce the quantum communication overhead; Kashefi and Wallden {{cite:a816c182b3216e027deed28af13778c8db5911ce}} also show how to reduce the overhead of {{cite:1eb91c53a6bbc1e4540227d9d400388fe4a2f4c7}}.",
"To the best of our knowledge, the proof techniques in these prior works appear as sketches only, or are cumbersome. In particular, the approach that uses quantum authentication codes {{cite:770d52cf7db37edfa831340270af1783aa0a5fa3}} is based on {{cite:b828601935cee1dea0d97c0cdf294fcf4a243fbd}}. However, the full proof of security for {{cite:b828601935cee1dea0d97c0cdf294fcf4a243fbd}} never appeared. Although {{cite:770d52cf7db37edfa831340270af1783aa0a5fa3}} makes significant progress towards closing this gap, it provides only a sketch of how the soundness is established in the interactive case. A full proof of soundness for {{cite:770d52cf7db37edfa831340270af1783aa0a5fa3}} follows from {{cite:e840756d50c5b53f74ab3cc2b96c8cc14131bed2}}, however the proof is very elaborate and phrased in terms of a rather different cryptographic task (called “quantum one-time programs”).In terms of the measurement-based approach, note that a proposed protocol for verification in {{cite:17837d66b991ff739fb81d2dc1700bc9750f4311}} was deemed incomplete {{cite:1eb91c53a6bbc1e4540227d9d400388fe4a2f4c7}}, but any gaps were addressed in {{cite:1eb91c53a6bbc1e4540227d9d400388fe4a2f4c7}}. In this case, however, the protocol (and proof) are very elaborate, and to the best of our knowledge, remain unpublished.",
"In sharp contrast to these approaches, Reichardt, Unger and Vazirani {{cite:03c13f9c98a9076d81787f5c5e334e8e4362e3b3}} show that it is possible to make the verifier completely classical, as long as we postulate two non-communicating entangled provers (this could be enforced, for instance, by space-like separation such that communication between the provers would be forbidden by the limit on the speed of light). The main technique used is a rigidity theorem which, provided that the provers pass a certain number of tests, gives the verifier a tight classical control on the quantum provers."
] |
[] |
Quantitative affine approximation for UMD targets
|
It is shown here that if $(Y,\|\cdot\|_Y)$ is a Banach space in whichmartingale differences are unconditional (a UMD Banach space) then there exists$c=c(Y)\in (0,\infty)$ with the following property. For every $n\in \mathbb{N}$and $\varepsilon\in (0,1/2]$, if $(X,\|\cdot\|_X)$ is an $n$-dimensional normedspace with unit ball $B_X$ and $f:B_X\to Y$ is a $1$-Lipschitz function thenthere exists an affine mapping $\Lambda:X\to Y$ and a sub-ball $B^*=y+\rhoB_X\subseteq B_X$ of radius $\rho\ge \exp(-(1/\varepsilon)^{cn})$ such that$\|f(x)-\Lambda(x)\|_Y\le \varepsilon \rho$ for all $x\in B^*$. This estimateon the macroscopic scale of affine approximability of vector-valued Lipschitzfunctions is an asymptotic improvement (as $n\to \infty$) over the bestpreviously known bound even when $X$ is $\mathbb{R}^n$ equipped with theEuclidean norm and $Y$ is a Hilbert space.
|
[
"Over the past several decades, research on quantitative differentiation has proceeded roughly along two lines of inquiry, one of which arising from functional analysis and metric geometry and the other arising from rectifiability questions in harmonic analysis. The present work belongs to the former direction, but its main contribution is the use of methods from the latter direction in this new context while incorporating various Banach space theoretic tools.",
"Bates, Johnson, Lindenstrauss, Preiss and Schechtman studied {{cite:696e8fb3ae1d6745b055914c896838f9ee247eb0}} quantitative differentiation in order to prove the rigidity of certain classes of Banach spaces under nonlinear quotients. The same notion was used in {{cite:0008f9f84ddcd63f2712f877d526ffb50fb9fcfb}} for metric embeddings, namely as an alternative approach to Bourgain's discretization problem {{cite:a893e6e8f3354a27c1d17fefeea0f48d44996a6b}} when the target Banach space is uniformly convex. The methods in this context fall under the category of (extensions of) “approximate midpoint arguments,\" as initiated by Enflo {{cite:9a20e5fb6dfaf2e3b4c129af33b3de1af01464a9}} to prove that {{formula:d5024c23-e7ee-4c04-b965-88bbf0a046e9}} is not uniformly homeomorphic to {{formula:fc92a379-48fb-42fb-b355-d67d422cf843}} , and further developed in {{cite:a893e6e8f3354a27c1d17fefeea0f48d44996a6b}}, {{cite:3f14512b5b4620803beb2c32c24478dad59faa25}}, {{cite:d6721f00085f00fa8077688fd2796ccc21e3fefa}}, {{cite:432f1a7f38d4f8cfac32e6b304923d647e4e48bc}} (see also Chapter 10 of {{cite:064441547d44c4a69b9507f438cbb181bc380aae}}). As examples of the many related applications to quantitative embedding theory and rigidity questions, see also {{cite:c84d33c507be957806bd28ba804a66add3507c56}}, {{cite:d14be8e696c8ffea1930411d71cec541739acb57}}, {{cite:3c3db2fe1d7d922d0ba2a13bd6c818fbe05c66a7}}, {{cite:3f14512b5b4620803beb2c32c24478dad59faa25}}, {{cite:51ec2af8dadc2cfaf7097fe0029c4696c4b94fa6}}, {{cite:5dcb81b3a9c6e13f7a09c7e2cc9f543b3fc0eaff}}, {{cite:994e57453186f961fc9014ab086d26250dc0f24f}}, {{cite:8befb8027f337a99af0d9ccc1bced2f3a8d6a1e6}}.",
"Parallel developments of a different nature arose in harmonic analysis, as part of the quest to develop a quantitative theory of rectifiability, with applications to singular integrals. Notable contributions along these lines include classical works of Stein (see the monograph {{cite:65fc5576796cdfd6caf6c48ec7daeca85b918160}}), through the works of Dorronsoro {{cite:e799a4603af6b6ce41b7059ef59a15c6117c0490}}, Jones {{cite:c5bc6b1f01b957e5d7e15316306251ad78a7eb52}}, {{cite:9be51b3b720ef2f0f8ab2f9338e2030a077ce32f}}, David and Semmes {{cite:e1f56c4551b157bfe20dd9b091b40f5238f72ede}}, {{cite:0a379ed67dd41315c31bf2c18925fe19378902b9}}, {{cite:100dc304da7761f5cd6563c01d46fec2df75523b}}, as well as the more recent work of Azzam and Schul {{cite:534147795285991496bc996de21d0e67c7eaa6b1}}. The work of Dorronsoro {{cite:e799a4603af6b6ce41b7059ef59a15c6117c0490}} directly influenced the present article, and we were also greatly inspired by the works of David and Semmes {{cite:e1f56c4551b157bfe20dd9b091b40f5238f72ede}}, {{cite:0a379ed67dd41315c31bf2c18925fe19378902b9}}, {{cite:100dc304da7761f5cd6563c01d46fec2df75523b}}. These works introduced and studied quantities that correspond to Definition REF when {{formula:26888970-8fa2-4b08-9993-6d02cd5fc51a}} is the Euclidean space {{formula:1f551ac2-51ce-4928-93fb-9bdd7870456f}} and {{formula:ccd6afa3-78ae-4a01-bde8-1195b588e5fb}} is the real line. Such methods also yield results for mappings from {{formula:2f4a169d-d57c-4e97-9d0b-829b1ec47639}} to {{formula:43d86162-d906-4769-b856-f652070ac6c1}} , but with statements that include implicit parameters that are allowed to depend on {{formula:378ff2f2-19b6-45b3-aa85-0d4b879b2010}} . In {{cite:100dc304da7761f5cd6563c01d46fec2df75523b}} David and Semmes compare their work to that of Bates, Johnson, Lindenstrauss, Preiss and Schechtman {{cite:696e8fb3ae1d6745b055914c896838f9ee247eb0}}, noting that the latter methods are different, and even yield results for infinite dimensional spaces.",
"Our contribution here follows the harmonic-analytic methodology, while overcoming several difficulties. Firstly, the literature on quantitative rectifiability ignores the dependence on the dimension {{formula:1ae00a24-bbe4-411e-848b-7608b6045135}} , while this dependence is the main topic of interest in the present context. In fact, a direct examination of the dependence on {{formula:cfc0e7f8-85eb-4912-b571-443f9e485f64}} that is implicit in the above cited works reveals that it is insufficient for the purpose of obtaining improved bounds on {{formula:e8ee54bc-a256-4edf-bbba-a186bbd7325f}} . Secondly, in our setting the domain {{formula:72ad9b76-c6a9-49c5-a191-1c2b672e9ea9}} is a general {{formula:b391167c-0889-4ea2-9ecc-c9c9034b1a42}} -dimensional normed space {{formula:cd6ebac2-bf54-4925-87fe-2b87e7badb50}} rather than a Euclidean space, and our arguments address this point. A final important difference is that we treat infinite dimensional Banach spaces {{formula:cab408eb-d6d6-40a1-9f06-2c26a0c15478}} as targets. Overcoming this requires substantial effort, because the infinite dimensional arguments of Bates, Johnson, Lindenstrauss, Preiss and Schechtman {{cite:696e8fb3ae1d6745b055914c896838f9ee247eb0}} do not seem to be applicable in our setting. The present work yields an infinite dimensional version of an inequality that Dorronsoro obtained {{cite:e799a4603af6b6ce41b7059ef59a15c6117c0490}} for real-valued functions. Such an infinite dimensional extension of Dorronsoro's work is not routine, and in particular it does not hold true for arbitrary infinite dimensional Banach space targets {{formula:98f01abb-0023-4b1d-a2ff-24ae37b5a8af}} . In fact, the geometry of {{formula:c8970125-abdc-4643-bcb6-b537fe9448da}} influences the structure of the inequality thus obtained, while stronger inequalities hold true for real-valued functions. The assumption that {{formula:0d709914-2245-44a4-9272-96f9f05a92ec}} is UMD is used several times in our argument through a rich UMD-valued Fourier-analytic toolkit that has been developed by many authors over the past four decades."
] |
[] |
Assisting Composition of Email Responses: a Topic Prediction Approach
|
We propose an approach for helping agents compose email replies to customerrequests. To enable that, we use LDA to extract latent topics from a collectionof email exchanges. We then use these latent topics to label our data,obtaining a so-called "silver standard" topic labelling. We exploit thislabelled set to train a classifier to: (i) predict the topic distribution ofthe entire agent's email response, based on features of the customer's email;and (ii) predict the topic distribution of the next sentence in the agent'sreply, based on the customer's email features and on features of the agent'scurrent sentence. The experimental results on a large email collection from acontact center in the tele- com domain show that the proposed ap- proach iseffective in predicting the best topic of the agent's next sentence. In 80% ofthe cases, the correct topic is present among the top five recommended topics(out of fifty possible ones). This shows the potential of this method to beapplied in an interactive setting, where the agent is presented a small list oflikely topics to choose from for the next sentence.
|
[
"Most works addressing “email text analytics” in the past few years have been on classification and summarization.Email classification has proven to be useful in many standard applications such as spam detection and filtering high-priority messages.Research themes such as summarization and question answering came into focus because of the need of better interpreting the overwhelming amount of emails/messages generated with the advent of email groups and discussion forums.One of the earlier contributions to email summarization was the work by {{cite:7a209f409f26d8e47a2fcdc4e182af485ccdfffc}} whereas {{cite:4e38af18522e21cfaed68fad270a0ada414d44c3}} extended it to email threads; Scheffer et al. Scheffer:04 on the other hand proposed semi-supervised classification techniques for question answering in the context of such threads.",
"Some of the recent classification and summarization techniques have been based on“speech acts” or “dialog acts” such as proposing a meeting, requesting information {{cite:ff41ba8a740cc9411e4bed4513883b80f744d5f4}}, {{cite:66a219c1a5dcfa5caa1a66627dff56d85f3536df}}. Several email studies including summarizing of email threads {{cite:65a15d0d17b6719c4752e3d557e78d729a93b7a6}} or classification of emails {{cite:9081a9bde2b38e1e73b1b5351f31d26fa729d7fd}} involve dialog-act based analysis.There has been very little work so far on customer-agent related email threads. Some of these works include identification of emotional emails related to customer dissatisfaction/frustration {{cite:74363483efebcab7622e740760dd83cffa1762f9}}, as well as learning possible patterns/phrases for textual re-use in email responses {{cite:76a56d574da6dedc1b1f83796c3aadd638242d0f}}. {{cite:e1fad16ddf99644beffdc143a56a182c4302313f}} is a recent work that attempts to generate emails based on a two-stage process where a structural template is first produced and then a topic-specific language model is used for producing textual realizations of the different slots in the template (see also {{cite:d7514cd9ace9ffc16851b713824c8d4c4daa308b}} for an earlier work using a similar language model based approach).",
"There have been a number of works that have addressed the problem of discovering the latent structure of topics in the related area of spoken and chat conversations. Recently {{cite:bae973a4167e4d9c5c5c89f18c3273bf5718f970}} have addressed this problem using HMM-based methods handling dialog state transitions and topic content simultaneously; this work differs from ours in several respects. First, the nature of the data is not the same, short alternating conversational utterances in their case, large single email responses in ours. Second, the focus of {{cite:bae973a4167e4d9c5c5c89f18c3273bf5718f970}} is on the discovery of latent topics (and conversational states) based on existing dialog texts (speech transcripts or chats), using HMM-based techniques different from our LDA approach. Finally, in our case, in addition to discovering the latent structure of existing emails, we also actually predict which topics will likely be employed in the forthcoming agent's response, which is not attempted in their paper."
] |
[] |
Fair and Square: Cake-Cutting in Two Dimensions
|
We consider the classic problem of fairly dividing a heterogeneous good("cake") among several agents with different valuations. Classic cake-cuttingprocedures either allocate each agent a collection of disconnected pieces, orassume that the cake is a one-dimensional interval. In practice, however, thetwo-dimensional shape of the allotted pieces is important. In particular, whenbuilding a house or designing an advertisement in printed or electronic media,squares are more usable than long and narrow rectangles. We thus introduce andstudy the problem of fair two-dimensional division wherein the allotted piecesmust be of some restricted two-dimensional geometric shape(s), particularlysquares and fat rectangles. Adding such geometric constraints re-opens mostquestions and challenges related to cake-cutting. Indeed, even the mostelementary fairness criterion --- proportionality --- can no longer beguaranteed. In this paper we thus examine the level of proportionality that canbe guaranteed, providing both impossibility results and constructive divisionprocedures.
|
[
"The most prominent geometric constraint in cake-cutting is one-dimensional:the pieces must be contiguous intervals. Several authors studied acircular cake {{cite:d17153c08496a75be52fe72dbf51748148b1b6c4}}, {{cite:b59c76433b32d6c3a181dc9a12c270a033545110}}, {{cite:aaaf1d61ebc9ef7c17a6d25e841b253a296ce3d2}},but it is still a one-dimensional circle and the pieces are one-dimensionalarcs.",
"Only few cake-cutting papers explicitly consider a two-dimensionalcake. Two of them discuss the problem of dividing a disputed territorybetween several bordering countries, with the constraint that eachcountry should get a piece adjacent to its current border: {{cite:55680d84a673cf4baf1581e8cbe7a29bcfaa04ee}}proved that such a partition exists and {{cite:748a99e9942f5691f2bbdf4f64da9cc84ef8e622}}complemented this proof with a division procedure. {{cite:122205a337c3fec37e755b0bfe1bfe3c36bfed6a}}describe a procedure that asks each of the {{formula:25b982b9-b0e9-4fef-8602-07931cf86bdd}} agents to draw {{formula:62e4e7f1-1b00-4201-8d0a-ad2ab7759d4c}} disjoint rectangles on the map of the two-dimensional cake. Theserectangles are supposed to represent the “desired areas” of theagent. The procedure tries to give each agent one of his {{formula:9178a3aa-4bc9-430a-bf2a-3752507da0b2}} desiredareas. However, the procedure does not succeed unless each rectangleproposed by an individual intersects at most one other rectangle drawnby any other agent. If even a single rectangle of Alice intersectstwo rectangles of George (for example), then the procedure fails andno agent receives any piece.",
"{{cite:4285e3145420aec34c4bea9df01a3bd93c9ebb83}}, {{cite:c3e7d19e63c6e424e0495b84789424193bac1e55}} acknowledge theimportance of having nicely-shaped pieces in resolving land disputes.They prove that, if the cake is a simplex in any number of dimensions,then there exists an envy-free and proportional partition of the cakeinto polytopes. However, this proof is purely existential when thecake has two or more dimensions. Additionally, there are no restrictionson the fatness of the allocated polytopes and apparently these canbe arbitrarily thin triangles. {{cite:654c9862d2178bab2e24db688268a16338c7ac9a}} studiesthe existence of competitive equilibrium with utility functions thatmay depend on geometric shape; their nonwasteful-partitions-assumptionexplicitly excludes fat shapes such as squares. {{cite:9b4a4388492d691966396e75db7d59ccc926c8b1}}studies a two-dimensional division problem in which the geometricconstraints are connectivity, simple-connectivity and convexity.",
"In our model (see Section ), the utility functionsdepend on gemoetry, which makes them non-additive. They are not evensubadditive like in the models of {{cite:492247a4fbad35bbad32081a505f68f3803c83dd}}, {{cite:8809fbfb48daec8bb5ee998c2589fc12afc00b05}}.Previous papers about cake-cutting with non-additive utilities canbe roughly divided to two kinds: some {{cite:654c9862d2178bab2e24db688268a16338c7ac9a}}, {{cite:dca881907917f66dd523f6451462644eb709f442}}, {{cite:c3e7d19e63c6e424e0495b84789424193bac1e55}}, {{cite:57a85621ad7023d40739b04c072163b57b86dd86}}handle general non-additive utilities but provide only pure existenceresults. Others {{cite:d2004125a38c2f0200dffa5535e531b97e870fa3}}, {{cite:3d69348fb9765c6ae9b7a4e1429733f1baae4610}}, {{cite:66d18996273478f360e52274e9c73d620ecc8468}}provide constructive division procedures but only for a 1-dimensionalcake. Our approach is a middle ground between these extremes. Ourutility functions are more general than the 1-dimensional model butless general than the arbitrary utility model; for this class of utilityfunctions, we provide both existence results and constructive divisionprocedures."
] |
[] |
Counting Euler Tours in Undirected Bounded Treewidth Graphs
|
We show that counting Euler tours in undirected bounded tree-width graphs istractable even in parallel - by proving a $\#SAC^1$ upper bound. This is instark contrast to #P-completeness of the same problem in general graphs. Our main technical contribution is to show how (an instance of) dynamicprogramming on bounded \emph{clique-width} graphs can be performed efficientlyin parallel. Thus we show that the sequential result of Espelage, Gurski andWanke for efficiently computing Hamiltonian paths in bounded clique-widthgraphs can be adapted in the parallel setting to count the number ofHamiltonian paths which in turn is a tool for counting the number of Eulertours in bounded tree-width graphs. Our technique also yields parallelalgorithms for counting longest paths and bipartite perfect matchings inbounded-clique width graphs. While establishing that counting Euler tours in bounded tree-width graphs canbe computed by non-uniform monotone arithmetic circuits of polynomial degree(which characterize $\#SAC^1$) is relatively easy, establishing a uniform$\#SAC^1$ bound needs a careful use of polynomial interpolation.
|
[
"Chebolu, Cryan, Martin have given a polynomial time algorithm forcounting Euler tours in undirected series-parallel graphs {{cite:ff06f7168115aee880e6657694fb2b7cf59319b8}} andthey have claimed to extend it to a polynomial time algorithm {{cite:5a88718ca97509832d64fb000c6d45b8384694e1}} for thecounting Euler tours in bounded tree-width graphs. We would like to point out thatthe only incomplete, unrefereed manuscript available publicly {{cite:5a88718ca97509832d64fb000c6d45b8384694e1}}sketches an algorithm that does dynamic programming directly on the tree-decomposition.Since we show how to obtain the line graph of the bounded tree-width graph efficientlyin parallel and then work on this bounded clique-width graph - our approachis fundamentally different from that of {{cite:ff06f7168115aee880e6657694fb2b7cf59319b8}}, {{cite:5a88718ca97509832d64fb000c6d45b8384694e1}}. Another differenceis that their algorithm is not designed to be parallelisable.",
"Also notice that in a precursor to this paper {{cite:d17580b4a3d026428f5f3b22ff16e6f65ae72036}},using totally different techniques (basicallyapplications of the Logspace version of Courcelle's theorem {{cite:d1b99f7dd77125e164fa5f69875e70842b8cc553}})it was claimed that the number of Euler tours in bounded tree-widthdirected and undirected graphs can be counted in Logspace but the approach hada serious flaw in the undirected version. Later versions {{cite:4529644030da0696b1d3e6a935ca99dfc16c2837}}, {{cite:a55420bc1466f9cfe5db5fac0a83a56fd2c8d475}}of the paper claim the result only for directed graphs. This work proves aslightly weaker version of the result - the upper bound being #SAC{{formula:641f5300-a6b8-494a-bb7b-691afa081131}} ratherthan Logspace.",
"Given that counting Hamiltonian cycles on bounded clique-width graphs willsuffice for our purposes, one result that is directly relevant is that ofFlarup and Lyaudet {{cite:2c7a70987d89303018022eff022107a58e552039}}: They study the expressive power of PerfectMatching and Hamiltonian polynomials of graphs of bounded clique-widthand show that they can simulate arithmetic polynomials, and arethemselves contained in {{formula:67e2c85c-02eb-4e05-9611-7d28ad51434a}} . This yields a {{formula:a49b4aab-621b-460e-848a-53d69210fcdd}} bound (implicit) for countingHamiltonian cycles in bounded clique-width graphs right away. There are twoaspects in which the work of {{cite:2c7a70987d89303018022eff022107a58e552039}} differs from our work: Firstly,even though their techniques are also inspired from {{cite:a9bda3b0785b5d9288758a0ff4b713540e7e0223}} like ours,they work with a slightly different notion of clique-width namely{{formula:f0b8bd31-d489-4805-9f3a-bde5476efbae}}These are weighted versions of clique-widthand are used to produce weighted graphs. {{cite:2c7a70987d89303018022eff022107a58e552039}} motivate this variant ofclique-width by observing that since {{formula:2fcd7b77-b57e-4433-add4-f98c255d6f0a}} has clique-width 2, most graphpolynomials are {{formula:1ae654ab-1e7f-484a-a102-f3f2c2895434}} -complete for bounded clique-width graphs.. Secondly,in the case of counting Euler tours, from a straight-forward application of {{cite:2c7a70987d89303018022eff022107a58e552039}}the best upper bound that can be obtained from the circuitfamilies constructed in {{cite:2c7a70987d89303018022eff022107a58e552039}} is non-uniform {{formula:21cc14dc-7ec9-4b04-8ae1-b61e253b5aa4}} , whereas we get an upper bound ofLogspace-uniformIn an earlier version of this paper, we had erroneouslyclaimed a GapL upper bound for counting Euler tours. As pointed out to us by Ramprasad Saptharishi,there is a rather serious gap with this approach. {{formula:e10cc379-1aa7-4d04-a7a4-ddfd594a2993}} .",
"There is some similarity that this work bears with that of Makowsky et al. {{cite:6e5682a6363023a5c4a2feba414e421f09d73de5}}, in thatboth involve polynomial interpolation to count witnesses for a graph theory problem. Thesimilarity is somewhat superficial because we use interpolation to obtain numbersindependent of the input graph while they interpolate to compute a graph polynomialthat crucially depends on the graph. The choice of graph theory problems is also quite different.In particular, {{cite:6e5682a6363023a5c4a2feba414e421f09d73de5}} does not address the Hamiltonian cycle problem."
] |
[] |
Towards Cleaning-up Open Data Portals: A Metadata Reconciliation
Approach
|
This paper presents an approach for metadata reconciliation, curation andlinking for Open Governamental Data Portals (ODPs). ODPs have been lately thestandard solution for governments willing to put their public data availablefor the society. Portal managers use several types of metadata to organize thedatasets, one of the most important ones being the tags. However, the taggingprocess is subject to many problems, such as synonyms, ambiguity orincoherence, among others. As our empiric analysis of ODPs shows, these issuesare currently prevalent in most ODPs and effectively hinders the reuse of OpenData. In order to address these problems, we develop and implement an approachfor tag reconciliation in Open Data Portals, encompassing local actions relatedto individual portals, and global actions for adding a semantic metadata layerabove individual portals. The local part aims to enhance the quality of tags ina single portal, and the global part is meant to interlink ODPs by establishingrelations between tags.
|
[
"The vast majority of scientific works about tagging and semantics focus on a different kind of context in relation to ours.Grubbers seminal paper {{cite:f9aef488789c81a108ee26c0f28aad3a7c382a59}}, and others such as {{cite:c53d47b81428f1868f85dbf26e3daec38a636152}}, {{cite:45a64f1e044f289587f2556f5fb96b8732259c97}}, {{cite:f8996d7be4fb6508d1d4ceb2e9073a52122b2c32}}, {{cite:594a364b9726db99ecd09d97922c59c62e47eff5}}, {{cite:cd9adeef9bf9799a335a02bd53b5e75eac772ce6}}, {{cite:64020edad5fa223b70f4c55923541774ee269054}} give interesting perspectives about the tagging activity and its relation to semantics, but always in the folksonomy (or collaborative tagging) context.In this case, tags are attributed to resources by the crowd, and thus, their quality are enhanced by crowd-selection mechanisms.This is applicable to platforms such as del.icio.us of flicker, where several users can tag the same resource.However, in the open data context tags are only attributed by system managers.The tag server approach described in this paper adds collaborative reconciliation layer over the ODPs.",
"In relation to the metrics for tagging environments, some related ideas could be found in the literature.For example, {{cite:0df726dfeceb36bb01f5c2c2e87c4bccd097647e}} presents a framework to evaluate the quality of ODPs.Among the applied quality metrics, three of them – Usage, Completeness and Accuracy – are related to metadata keys, which tags are part of.Usage establishes which metadata keys are actually used in a portal; Completeness evaluates the presence of non empty values; and Accuracy checks if metadata adequately describes the data.However, this metric is not applied for tags.",
"Laniado and Mika did a similar analysis over hashtags on Twitter {{cite:6c96a4fcd1d450b328bf3db061058d6be82ceac6}}.Their work is focused in answering if Twitter hashtags constitute strong identifiers for the semantic web.To achieve this, four metrics are used: frequency of hashtags; specificity, which is the deviation from the use of them without being a hashtag; consistency; and stability over time.",
"The problem of semantic lifting in ODPs was tackled by {{cite:f206d47292cdf069e1c89dccec0ea81b11a03fe2}}, {{cite:68bf5e91dbb8f97d08ad791f92b87e367be77eef}}.In {{cite:e24e8dbecee969b122806bdbc835057a0bf924e9}}, a strategy for lifting datasets in ODPs to the Linked Data cloud is presented.In all these works, however, the semantic lifting refers to the datasets, and not to metadata.",
"There also has been some work done with regard to metadata reconciliation {{cite:00c22b0313e70a93e301b56162bca6f08aa9d33c}}, {{cite:1b960b556cb599053c7b0be2fd3c15945f29ce7c}}.However, to the best of our knowledge none of them has been specifically applied to open data portals or leverages tag curation as proposed by STODaP."
] |
[] |
A Graph Traversal Based Approach to Answer Non-Aggregation Questions
Over DBpedia
|
We present a question answering system over DBpedia, filling the gap betweenuser information needs expressed in natural language and a structured queryinterface expressed in SPARQL over the underlying knowledge base (KB). Giventhe KB, our goal is to comprehend a natural language query and providecorresponding accurate answers. Focusing on solving the non-aggregationquestions, in this paper, we construct a subgraph of the knowledge base fromthe detected entities and propose a graph traversal method to solve both thesemantic item mapping problem and the disambiguation problem in a joint way.Compared with existing work, we simplify the process of query intentionunderstanding and pay more attention to the answer path ranking. We evaluateour method on a non-aggregation question dataset and further on a completedataset. Experimental results show that our method achieves best performancecompared with several state-of-the-art systems.
|
[
"Question answering (QA), which is to return the exact answers to a given natural language question, is a challenging task and has been advocated as a key problem for advancing web search. Previous work is mainly dominated by keyword based approaches, while recent blossom of large-scale knowledge bases have enable numerous KB-based systems. A KB-based QA system answers a question by directly querying structured knowledge knowledge, which can be retrieved using a structured query engine.",
"Recently many works have been published in this field.Apart from {{cite:f6de934ebf0cae312234fc395acd2582dbba02a3}}, {{cite:cc17e2eb2699857165493c1cc8011a058c3feb68}}, {{cite:543e28f5074da0d118d730cd01ae147242454b7b}} discussed in Section ,PowerAqua {{cite:459a5cbff2b66f65b57f836dc87b565b2d52c2b5}} proposes a natural language user interface making people query and explore semantic web content more convenient.Two research work {{cite:2119f0534510b2fd769689adc02eee11835c4e2c}}, {{cite:de848e831438e418b82439739c7bbc6166bcaf86}} present an ILP(Integer Linear Programming) method to translate a natural language question into a structured SPARQL query.The Paralex system {{cite:d92c92ca5b51073bca2ca2ed04451f5e0483309d}} studies question answering as a machine learning problem and induces a function that maps open-domain questions to queries over a database of web extractions.Meanwhile, Shekarpour et al. {{cite:68c2777166e2875029da84c3a21b50cb94953269}} presents an approach for question answering over a set of interlinked data sources. This approach firstly employs a Hidden Markov Model to determine the most suitable resources for a use-supplied query and secondly constructs a federated formal query using the disambiguated resources and linking structure of underlying datasets.And two research work {{cite:ae41c1267bd37caa2a336367d3bf583ddd82dd30}}, {{cite:017a92d433491db94f4f1257a897cf5a298d72b8}} develop semantic parsing techniques that map natural language utterances into logical form queries, which can be executed on a knowledge base.Xu et al. {{cite:443f094876a2e8853ced2ceee18c4c65f4534b14}} develops a transition-based parsing model to do semantic parsing for aggregation questions.",
"However, most of these methods focus on translating a question to a SPARQL query. Meanwhile many methods need many well designed manual rules. Also lots of them focus on recognizing the inherent structure of user's query intention using different semantic parsing techniques implemented by complex models. In contrast, we aim to finding the most appropriate path rather than generate SPARQL query templates directly. And we simplify the process of query intention understanding and pay more attention to the answer path ranking. Our approach is simple in structure but effective in terms of performance."
] |
[] |
Learning multi-faceted representations of individuals from heterogeneous
evidence using neural networks
|
Inferring latent attributes of people online is an important social computingtask, but requires integrating the many heterogeneous sources of informationavailable on the web. We propose learning individual representations of peopleusing neural nets to integrate rich linguistic and network evidence gatheredfrom social media. The algorithm is able to combine diverse cues, such as thetext a person writes, their attributes (e.g. gender, employer, education,location) and social relations to other people. We show that by integratingboth textual and network evidence, these representations offer improvedperformance at four important tasks in social media inference on Twitter:predicting (1) gender, (2) occupation, (3) location, and (4) friendships forusers. Our approach scales to large datasets and the learned representationscan be used as general features in and have the potential to benefit a largenumber of downstream tasks including link prediction, community detection, orprobabilistic reasoning over social networks.
|
[
"Much work has been devoted to automatic user attribute inference given social signals.For example, {{cite:03047a1840754a270b1da24c5194c58452459479}}, {{cite:9b5c03c76273f7afbb9bc3f4d33fed98a24c07e7}}, {{cite:ffee04f9549fe876ceb91c5fd8c7d05ba7296dbd}}, {{cite:29368c8809e5fab5e1b8acf8b84b7fcd4a82e2ca}}, {{cite:2b0bbb8025669c7aafb3a6d172bb381b7284b5c9}}focus on how to infer individual user attributes such as age, gender, political polarity, locations, occupation, educational information (e.g., major, year of matriculation) given user-generated contents or network information.",
"Taking advantage of large scale user information, recent research has begun exploring logical reasoningover the social network (e.g., what's the probability that a New York City resident is a fan of the New York Knicks).Some work {{cite:e4d8244976916a7d8978eba96bf676b6c0e24a1d}}, {{cite:a3a7ab32c423416395ba4cd3483abc5862b05854}} relies on logic reasoning paradigms such as Markov Logic Networks (MLNs) {{cite:5dfbba083417944f5d096176c5d5331a5a42e6be}}.",
"Social network inference usually takes advantage of the fundamentalpropoety of homophily {{cite:1d515d72832b3ffa79b6d82183887569fe4883db}}, which states that people sharingsimilar attributes have a higher chance of becoming friendsSummarized by the proverb “birds of a featherflock together\" {{cite:7be572d372d21ee3d1d9b75b216da4030ba28f8f}}., and converselyfriends (or couples, or people living in the same location) tend to share more attributes.Such properties have been harnessed for applications like community detection {{cite:95717a3fc7e980896e5558ae51c12d442115b891}} and user-link prediction {{cite:9f7f55433c412fc961c1fa4a5567d82227f3bb26}}, {{cite:bcca9658185cdf613cb7dfc36055e5e433f078d3}}.",
"The proposed framework also focuses on attribute inference, which can be reframed as relation identificationproblems, i.e., whether a relation holds between a user and an entity. This work is thus related toa great deal of recent researches onrelation inference (e.g.,{{cite:1c1e8f90c56825bb76c636a91cd7bb88c7e0f6ad}}, {{cite:41ef1d09b164972dcca6f483df8c9787e23bb4e1}}, {{cite:592b4d9f9a491d610f2763dbd349912973512150}}).",
"Our work is inspired by classic work on spectrallearning for graphse.g., {{cite:ce67af9d30b3d843e5b3caa0ca49ed551f4983f6}}, {{cite:42ae23359e5d0e260046ac8049ad6d390f86db6a}} and onrecent research {{cite:9f7f55433c412fc961c1fa4a5567d82227f3bb26}}, {{cite:ac431572dd0fa192d7dff641451422ed56834770}}thatlearn embedded representations for a graph's vertices. Our model extends this workby modeling not only user-user network graphs, but also incorporatingdiverse social signals including unstructured text, user attributes, and relations,enabling more sophisticated inferences and offering an integrated model of homophily in social relations."
] |
[] |
Shortest paths in one-counter systems
|
We show that any one-counter automaton with $n$ states, if its language isnon-empty, accepts some word of length at most $O(n^2)$. This closes the gapbetween the previously known upper bound of $O(n^3)$ and lower bound of$\Omega(n^2)$. More generally, we prove a tight upper bound on the length ofshortest paths between arbitrary configurations in one-counter transitionsystems (weaker bounds have previously appeared in the literature).
|
[
"Reachability is a fundamental problem in theoretical computer science and in its applications in verification,notably via analysis of infinite-state systems {{cite:bb710eafc01984ff2754fb5a8787068243f81003}}, {{cite:58a9caa118e7f86db81a2e5c0b4efd2842e4c5f2}}, {{cite:3c80bded8c00da26400c89241dd858fcadaff9b6}}, {{cite:47877aea8526be7dd8a07c9deca3440b863dade9}}.Among such systems, counter-based models of computation are a standard abstractionin many domains and have attracted a lot of attention {{cite:32e90d34ef7a802ad55b504d51d926b1924afe8f}};machines with a single counter are, of course, the most basic.Nevertheless, while our main motivation has been purely theoretical,we find it interesting to note that bounds on the length of shortest pathsin one-counter systems have appeared as building blocksin the literature on rather diverse topics.",
"More specifically,a polynomial upper bound is used by Etessami et al. {{cite:746a8ee93478c5acc1bea28d1d16b0be6c725c63}} in an analysis ofprobabilistic one-counter systems (which are equivalent to so-calleddiscrete-time quasi-birth-death processes, QBDs).They prove that in the {{formula:b07829f4-a534-491f-8d6b-89f205af99a0}} -reachability settingthe counter does not need to grow higher than {{formula:f51db1f7-c80c-4209-bdca-81ca58a76650}} and provide examplesshowing that this bound is tight.However, they only deduce upper bounds of {{formula:9a216335-5d8f-4c3d-a2b6-ba3b6549430f}} and {{formula:89640b04-6e38-46c3-a91f-6ef9714a9618}} on the length ofshortest paths without and with zero tests, respectively.A simple corollary shows that if a state {{formula:290f8824-b93b-46b8-8845-6a10d8f86cff}} can eventually reach a state {{formula:45b94893-4979-490e-9a8f-eb5e5ad95278}} with a non-zero probability, then this probability is lower-bounded by {{formula:35e6e204-d42d-42e8-a7ed-be2ee5bfdb55}} where {{formula:1ead296d-0b3c-469a-88d5-166f58f0fbef}} is the smallest among positive probabilities associated with transitions.This becomes a step in the proof that a (decomposed) Newton's methodapproximates termination probabilities of the systemin time polynomial in its size, {{formula:0e1208a2-b0c7-41ed-a202-8de7bc5e2f33}} ;the results of the present paper reduce the (theoretical) worst-caseupper bounds on the number of steps roughly by a factor of {{formula:2b41c89d-7883-4215-8802-e6b5f752fa53}} .",
"In a subsequent work, Hofman et al. {{cite:edc8bd769d03565889be9e6902ba08c98c71dc0e}} reuse the auxiliary lemmason the length of shortest paths from {{cite:746a8ee93478c5acc1bea28d1d16b0be6c725c63}}and show that (strong and weak) trace inclusion for a one-counter system anda finite-state process is decidable in PSPACE (and is, in fact, PSPACE-complete).",
"One may note that a stronger upper bound of {{formula:9404a2c8-fe97-4cb4-92dc-28f6a6a2d1aa}} on the length of shortest pathscan be derived from the above bound on the largest needed counter valueeven in the presence of zero tests.This value, {{formula:c99928f8-8110-45ee-a6d8-d4eb4efbf053}} , seems to be a recurring theme in the literature on one-counter systems;it already appears in the pumping lemma for one-counter languagesdue to Latteux {{cite:750769c964c5fa46a80fedb12c52b7d2460fe2e8}} as the pumping constant:a number {{formula:5f10abab-decd-42be-9df8-66331d9c9571}} such that any accepted word longer than {{formula:c32a2b38-2314-498a-a6dd-c3c38bcb731b}} can be pumped.In fact, the formulation in {{cite:750769c964c5fa46a80fedb12c52b7d2460fe2e8}} does not permit removalsof factors from an accepted word, but even such a version would only yieldthe same upper bound of {{formula:46d37aa7-bc3e-4a6c-b372-cd7cb364485d}} on the length of shortest paths.While the arguments of the present paper do not lead to an improvement inthe pumping constant for one-counter languages,we nevertheless show that in the reachability settingthe optimal value (the length of the shortest path) is actually {{formula:6eef4fe4-9f9e-4c2b-a4ae-545a8ea01c93}} .",
"A cubic upper bound on the largest needed counter value(for the reachability setting) in one-counter systemswithout zero tests, also known as one-counter nets, appears inthe work of Lafourcade et al. {{cite:7761a3a8ae4d0634b55fe20f67ef696afb9306cb}}, {{cite:d8787a905a0790cff388916ef99c927a1209a52a}}.This result is applied in the context of the Dolev-Yao intruder model,where the question of whether a passive eavesdropper (an intruder) can obtaina piece of information is reduced to the decision problemfor a deduction system.For several such systems, Lafourcade et al. show that,under certain assumptions,the problem is decidable in polynomial time.They construct a one-counter system wherestates represent terms from a finite set and the counter valuecorresponds to the number of applications of a free unary function symbolto a term. After this, the upper bound on counter values along shortest pathsis extended to an upper bound on the size of terms that can be usedin a minimal deductive proof;needless to say, an improvement in the upper bound extends in a natural way.",
"Finally, we would like to mention the work of Alur and Černý {{cite:0bdbe03138c0bfafbd635ef2976963f9bb05783e}},who use a related model of one-counter systems with counter values in {{formula:2ad46407-2d2b-4471-a4d4-7e89bfd9ddbc}} and without zero tests.They reduce the equivalence problem for so-called streaming data-stringtransducers to {{formula:887cba25-8af3-425e-9ef6-62a67086c460}} -reachability in such counter systems:the transducers produce output at the end of the computation, andthe counter is used to track the accumulated distance betweena distinguished pair of symbols in the output.Since these transducers are designed to model list-manipulating programs(in two syntactically restricted models),decision procedures for equivalence of such programscan rely on the upper bounds for shortest paths to efficiently prunethe search space.In {{cite:0bdbe03138c0bfafbd635ef2976963f9bb05783e}}, the upper bound on the path length is the familiar {{formula:2d00a37f-5ee0-474e-b2fe-cb21054f815e}} ;this gives an upper bound on the length of smallest counterexamples to equivalence.Our upper bound of {{formula:04045971-fc34-4a87-bc69-05661236fa2b}} extends to this model of counter systems too.The reduction to reachability in one-counter systemswas recently implemented by Thakkar et al. {{cite:75f95c5b6614860519f444bc919f203c8a27226a}}on top of ARMC, an abstraction-refinement model checker {{cite:3676a8c7cc4a5a79f3a83e84789662af73cd1a33}},for the purpose of verifying retransmission protocols over noisy channels."
] |
[] |
Towards Patterns for Heaps and Imperative Lambdas
|
In functional programming, point-free relation calculi have been fruitful forgeneral theories of program construction, but for specific applicationspointwise expressions can be more convenient and comprehensible. In imperativeprogramming, refinement calculi have been tied to pointwise expression in termsof state variables, with the curious exception of the ubiquitous but invisibleheap. To integrate pointwise with point-free, de Moor and Gibbons extendedlambda calculus with non-injective pattern matching interpreted usingrelations. This article gives a semantics of that language using ``idealrelations'' between partial orders, and a second semantics using predicatetransformers. The second semantics is motivated by its potential use withseparation algebra, for pattern matching in programs acting on the heap. Lawsincluding lax beta and eta are proved in these models and a number of openproblems are posed.
|
[
"It does not appear that the specific calculus of de Moor and Gibbons has been developed beyond the original paper {{cite:d825a1bae91e8210e3ac61f859ce0f37a0c1dad5}}.",
"Hinze makes nice use of explicit powersets as a way to do pointwise relational programming {{cite:bf06f09ba381c0e8af473e8750ac5f4f7ce5614f}}.",
"Bunkenburg {{cite:ed3484edfcf0f7b9c53d90ae2d65e4a53e113158}} develops a calculus of“expression refinement”, for derivation of lazy functional programs including those that use the state monad.Ordering is taken into account by treating data types asordered sets. Nondeterministic expressions are interpreted as monotonic functions{{formula:d52f7aaa-dea4-4cad-a823-bdb7f99be5d0}} where {{formula:18846e90-6bd2-4ef2-aa0c-d1a886a710c7}} is the set of updeals (upward closed subsets) on {{formula:e80c717e-f18b-44d3-9078-b99d8dc3ccdb}} .The set {{formula:87e23c98-a39a-4e84-aa21-bc5d85f9d48e}} is ordered by {{formula:6136de7b-f760-4515-8926-b586db97be1e}} , which coincides with the Smyth order(total correctness) restricted to upward closed sets. Upward closure is natural foroutcomes of expressions.Monotonic functions {{formula:2c7405f4-14f9-4670-82bd-9fee8397db75}} are naturally isomorphic to ideals, as isdiscussed in Section .So there is some overlap with the present work,although our language (and {{cite:d825a1bae91e8210e3ac61f859ce0f37a0c1dad5}}) is by-value.Bunkenburg develops a wide spectrum language and a logic of refinement, but the language does not include pattern matching.",
"Uses of angelic and demonic nondeterminacy are explored by Back and von Wright {{cite:773fbe2e3b26f43d2ed1e044c1c15569af85f373}} and by Morris and Tyrrell {{cite:3d5a06e15cf16f4a99de9c857f9c420c3ed36c98}}.Martin, Curtis, and Rewitzky {{cite:e5c5148ffba5e3dce38d4315d7e2a75fee1a66a0}} use multi-valued relations as an alternative to predicate transformers for modelling the combination of angelic and demonic nondeterminacy,and adapt map/fold algebra to that model {{cite:f3abe3fe2480bfa09400a951b97384ecded84fc1}}.",
"Morris, Bunkenberg, and Tyrrell {{cite:aab06175bb00f7327b2bf5045e110b3f521c26dc}}introduce a novel “phrase” construct, and “term transformer” semantics, as a technique for specifying and reasoning about imperative programs; it generalizes predicate transformers to terms of any type.Their approach encompasses unbounded angelic and demonic nondeterminacytogether with higher order functions, for which Morris and Tyrrell present numerouslaws {{cite:3d5a06e15cf16f4a99de9c857f9c420c3ed36c98}}.They sketch how non-injective patterns may be reduced to nondeterminacy.They have beta-reduction as an equality, unlike our results in the sequel,owing to partitioning terms as proper or not, and use of a semantic monotonicity restriction.Morris and Tyrrell have also shown equivalence between monotonic predicate transformers, multirelations, and free completely distributive lattices {{cite:e1915475b8708a70d4ca170181b7b01d245663b7}}.The phrase construct is very appealing but quite different from the functional notations explored in this article,and from conventional notations of refinement calculi.",
"For first order concurrent programs, concurrent Kleene algebra provides a pointfree algebraic setting for reasoning about separation, mutable state and concurrency (e.g., {{cite:98ec7ebadd34c504afb9adff827d5bc307582c14}}).Modal Kleene algebra is the basis for the extended separation logic of Dang and Möller {{cite:d58e5647094adf4e10d0f1e365813fc5490b88a5}},which they use to verify algorithms including list reversal and tree rotation, expressed in terms of load and store commands.Banerjee et al {{cite:44661809258eddacce68050ac6da02061be48cfa}} use explicit expressions to designate footprints of pointer structures, but only for post hoc verification."
] |
[] |
Discriminative Predicate Path Mining for Fact Checking in Knowledge
Graphs
|
Traditional fact checking by experts and analysts cannot keep pace with thevolume of newly created information. It is important and necessary, therefore,to enhance our ability to computationally determine whether some statement offact is true or false. We view this problem as a link-prediction task in aknowledge graph, and present a discriminative path-based method for factchecking in knowledge graphs that incorporates connectivity, type information,and predicate interactions. Given a statement S of the form (subject,predicate, object), for example, (Chicago, capitalOf, Illinois), our approachmines discriminative paths that alternatively define the generalized statement(U.S. city, predicate, U.S. state) and uses the mined rules to evaluate theveracity of statement S. We evaluate our approach by examining thousands ofclaims related to history, geography, biology, and politics using a public,million node knowledge graph extracted from Wikipedia and PubMedDB. Not onlydoes our approach significantly outperform related models, we also find thatthe discriminative predicate path model is easily interpretable and providessensible reasons for the final determination.
|
[
"Discriminative Path Generation. Although meta paths have been used in many methods such as similarity search, clustering, semi-supervised learning, and link prediction {{cite:12a02cc87991109eb070906879a0fc23716b85f7}}, {{cite:ce2ce15c4055699488d16ac2dfc926030a530055}}, {{cite:57ffda32eaae0071e7e8c8f6163c160afe362a6a}}, {{cite:15a82c01a4bc62fb648e797517617e871797a7fe}}, {{cite:bebd7b8768635b3e546dd922534934f6e42f8c3c}}, these algorithms either require human annotated meta paths {{cite:2d8bb3f21f7333dd1483c3056ac64fce09465b98}} or enumerate all possible meta paths in the graph. Recently efforts have been made to meta path discovery and association rule mining in concept graphs {{cite:112507792cc23315bcec4edca1446212f5487dad}}, but most of the approaches have their own limitations. Meng meaning:NTF . et al catcode:NTF a et al. et al., proposed a meta path generation algorithm that prunes the enumeration space by logistic regression, but this approach is prone to premature rejection and may miss important discriminative paths {{cite:bda8604f219b82ab448d7fddaba6460631ff54db}}. AMIE {{cite:9038dc2936d648845468e27dea5b6fdb6a125174}} is a global association rule mining algorithm which can not mine personalized, i.e., context dependent, association rules as we shown in Sec. and . Abedjan and Naumann proposed a predicate expansion algorithm {{cite:23094e029268280703f0dc1ffa960ba471e30117}} which can find predicate synonyms, but cannot find predicate paths that have discriminating power. The proposed discriminative path discovery framework in this work extracts meta paths and predicate paths from the graph directly with given endpoints, therefore our framework will not miss important predicate paths in the graph and is context-sensitive.",
"Fact Checking. With the large volume of data generated every day, the number of unverified statements begets the need for automated fact checking {{cite:29cf8e5c38db295bbd8d66a9ef07fca725410c1c}}, {{cite:795830d437abda69385e0da816ab9f456f931bcb}}. To that end, many researchers have focused on automated fact checking in recent years. Finn meaning:NTF . et al catcode:NTF a et al. et al. introduced a new interactive tool to help human fact checkers determine the quality of a statement by extracting the propagation of facts on Twitter {{cite:9bb19435e3a1c507573a434eb1710f8cc861aa7d}}. Ennals meaning:NTF . et al catcode:NTF a et al. et al. created a crowd-sourced platform that highlight disputed claims {{cite:bb149980f2d28569a5528d7b999db6c45b9c1b45}}. Kwok meaning:NTF . et al catcode:NTF a et al. et al. proposed a ensemble method utilizing the result from search engines to check a given statement {{cite:ba41839c29ae9eb100bf37529ac057d8efc2fee6}}. Hassan meaning:NTF . et al catcode:NTF a et al. et al. proposed a numerical fact monitor, called FactWatcher {{cite:581eb912d7b9a6512dca32daed86f26bf9aabb5a}}, that uses an append only database and certain skyline operators {{cite:8e6a17b9932fc39def6f6ae923ebdb12aae53ba8}}, {{cite:65fef9e8a497f026f4a4d1fb4afe3b371714862d}}, but FactWatcher is not applicable to knowledge graphs or nonnumerical statements. The True Knowledge System {{cite:09f330496fc46fec83ddadb354b7954dbe5709de}} validates a statement of the fact using {{formula:639d1a12-a376-483c-81c0-dead18c0f618}} predefined and user provided association rules; unfortunately, this means that it is impossible to check a statement that does not already have a predefined association rule within True Knowledge. Ciampaglia meaning:NTF . et al catcode:NTF a et al. et al. published a knowledge graph based fact checking algorithm {{cite:d3f41e0a5afc8e8865c0b66da7e6958e9a007727}} utilizing node connectivity, but does not take advantage of the type-labels in the heterogeneous information networks. Recently, Guu meaning:NTF . et al catcode:NTF a et al. et al. published a question answering algorithm that converts a given question into a vector space model to find the answer {{cite:57ae7fb148c097dcc4dbdb7ed240dcac02e122dd}}, but, like neural network based models {{cite:480dbdcfb6587446cfcbcaf6630b569106f5ac69}}, the learned model is generally uninterpretable. Li meaning:NTF . et al catcode:NTF a et al. et al. proposed T-verifier, a search engine based fact checker {{cite:28a87155959b89d258cf871d789aa85f58654366}}, but such approach needs extensive access to search engine APIs which is not easy to gain. Knowledge graph completion methods, such as TransE {{cite:2d5f38e66ed4d94ed8ecde1f8cb63315b4ab196a}}, TransR {{cite:ee4237419b7c680b09214a3c1b865ec703dd2762}}, and NTN {{cite:f5d1ba76f399dd334c4b69bbfdc0e7981c0b6e02}} is not ideal for fact checking in large knowledge graphs because the slow convergence rate.",
"Link Prediction. Apart from classic homogeneous link prediction methods, such as Adamic/Adar {{cite:6019967500d7d4d75af54e7309bb495bd7eb8081}}, SimRank {{cite:abc766c3ddc7aa462f5597ec0ba6309727cd9b78}}, Katz {{cite:1a420998238614c33dab1a634babdec5590c1e1a}}, Preferential Attachment {{cite:b4bd63e9be5ca5646de3244fa37690d783eb3621}}, and Personalized PageRank {{cite:4f11645ea299e9b6d4c51a969cae83ea736e944d}} etc., many heterogeneous methods have been developed to leverage the rich information in heterogeneous information networks. Heterogeneous graphlet base methods {{cite:de0f373c8e72633685db0d5bbff05746f1f2b952}} predict the relation between two endpoints by counting the occurrence of certain heterogeneous motifs, which are not applicable to complex knowledge graphs due to the exponential number of possible graph motifs. Other heterogeneous link prediction methods that adapt from classic homogeneous algorithms, HeteSim {{cite:12a02cc87991109eb070906879a0fc23716b85f7}} and PCRW {{cite:bebd7b8768635b3e546dd922534934f6e42f8c3c}}, depend on human annotated meta paths. PathSim {{cite:ce2ce15c4055699488d16ac2dfc926030a530055}}, a heterogeneous similarity metric, also requires hand crafted and symmetric meta paths as the input. Recently Dong meaning:NTF . et al catcode:NTF a et al. et al., proposed a heterogeneous link prediction algorithm based on coupled networks, but also needs human annotated meta paths as input {{cite:5cf8af4a9687e4914e7bfb47b020b43a371434c1}}. In contrast, this work automatically discovers the important meta paths and predicate paths that related to the given statement of fact."
] |
[] |
Exclusive Sparsity Norm Minimization with Random Groups via Cone
Projection
|
Many practical applications such as gene expression analysis, multi-tasklearning, image recognition, signal processing, and medical data analysispursue a sparse solution for the feature selection purpose and particularlyfavor the nonzeros \emph{evenly} distributed in different groups. The exclusivesparsity norm has been widely used to serve to this purpose. However, it stilllacks systematical studies for exclusive sparsity norm optimization. This paperoffers two main contributions from the optimization perspective: 1) We provideseveral efficient algorithms to solve exclusive sparsity norm minimization witheither smooth loss or hinge loss (non-smooth loss). All algorithms achieve theoptimal convergence rate $O(1/k^2)$ ($k$ is the iteration number). To the bestof our knowledge, this is the first time to guarantee such convergence rate forthe general exclusive sparsity norm minimization; 2) When the group informationis unavailable to define the exclusive sparsity norm, we propose to use therandom grouping scheme to construct groups and prove that if the number ofgroups is appropriately chosen, the nonzeros (true features) would be groupedin the ideal way with high probability. Empirical studies validate theefficiency of proposed algorithms, and the effectiveness of random groupingscheme on the proposed exclusive SVM formulation.
|
[
"Structural sparsity induced by regularization terms have been widely used recently for the feature selection purpose {{cite:606f0d630caeca54ba407a999a1dddff3488da70}}, {{cite:6108244103146b887a448849726ad95f6aa88241}}. Both theoretical and empirical studies have suggested the powerfulness of structure sparsity for feature learning, e.g., LASSO {{cite:7aebc74443becb544bc93310e9eeb239e4b5df18}}, group LASSO {{cite:d07d675b39b99bbbc9405f54416b188f59171be9}}, exclusive LASSO {{cite:7661f48f146d78b5ea7dca07e2165312da7938e8}}, fused LASSO {{cite:4b0d06c84fab87601f47457082f395a2d39b0617}}, {{cite:16461d3a0e097be57602e311b008cea319c20827}}, and generalized LASSO {{cite:b166cf745c9f58b5bc038809db77500ef96aa257}}. We refer readers to {{cite:606f0d630caeca54ba407a999a1dddff3488da70}} for comprehensive review.",
"The {{formula:5cb6617d-e92f-49c0-8fe5-9c826a4273fe}} norm minimization problem, particularly the LASSO formulation {{cite:7aebc74443becb544bc93310e9eeb239e4b5df18}}, has received extensive attentions. Many efficient algorithms have been proposed to solve it, including proximal gradient descent {{cite:03797004b91ea2f0e6be850d39f27293ba691e6f}}, accelerated gradient descent {{cite:0f51b3214e6f236d07fe7ac114e56b834bdc1a39}}, {{formula:6699f436-efe2-4c7d-96db-2447c1e3fedc}} {{cite:d18761f6f1889c14d30f89685872d2475742d4de}}, coordinate descent {{cite:5a838adf256b289f151500a8a6521d800af94509}}, accelerated coordinate descent {{cite:e9b4469d9568be611aee639f1f848bba19da92b2}}, and interior point method {{cite:1feeca4c1d566cef59773a3351533bce57652a79}}. We refer readers to {{cite:ff0acfba885f995fd92586fd850b94f9540bc2f0}} for comprehensive review of the {{formula:eb11dc55-ac79-424d-b089-c27d4c3e0ff0}} norm optimization.",
"The {{formula:987ad869-5b60-42d4-a459-ec90d07bc194}} (group sparsity) norm minimization, particularly the group LASSO (without overlaps) {{cite:d07d675b39b99bbbc9405f54416b188f59171be9}}, can be solved using similar optimization algorithms to {{formula:9d510087-657a-45da-87e4-90eb789ab914}} norm problems by simply treating each group as an element {{cite:c1c9a6d4689d8905db60249be1942b087058f9e5}}. For the the group LASSO with overlaps, one can use the {{formula:54ea2616-bc98-41c8-a24f-02f0dc7d373d}} framework to solve it {{cite:e39a461c3f3a82e2c1a764658dd0dc3b540df495}}, where the key step can still apply those approaches for non-overlapping formulations.",
"The exclusive sparsity regularization was originally defined on matrices for the feature selection purpose in multi-task learning {{cite:7661f48f146d78b5ea7dca07e2165312da7938e8}}. A subgradient algorithm is proposed to solve this problem, but leads to a slow convergence rate {{formula:d57a8475-a6f8-479e-a125-21e430dcc110}} , where {{formula:2684e3be-b971-4c2e-90f7-65cd0f5c1e86}} is the number of iterations. {{cite:f1ed2f6a8d6cc7c4252b872cd729d00a36a0a9dd}} considers to minimize an objective “{{formula:46ac96d0-29ca-4c54-8b6c-775fb791a3bc}} ” and proposes a reweighed approach to solve it. However, the proposed reweighted approach only guarantees convergence, which means that the convergent point is not necessarily the optimal solution. {{cite:c86ee2b3fbbd850b85185494f5a37f64085eeead}} considers to minimize an variant objective “{{formula:0b9821a4-418b-4db7-986a-a6d069e887c2}} ”. They apply the Nesterov's smoothing scheme to smooth the regularization term and the Nesterov's accelerated approach to solve an approximate smooth objective function {{cite:f439fed08bf1b777406ab75ae6eb8fdf29ebd592}}. This algorithm leads to a convergence rate {{formula:1cd9fff4-9d0a-4667-a42a-31b4bf27e6f1}} . To the best of our knowledge, none of existing algorithms can strictly achieve the optimal convergence rate {{formula:5d7975fd-848a-4bf3-b9c1-92d49f2caa87}} for the general exclusive sparsity norm minimization even the loss function is smooth.",
"[H] The General Framework of {{formula:613d7f6f-593e-458b-bfe6-75df05efde33}} [1] {{formula:2b825464-2db3-4512-a53e-e4134a45efc1}} , {{formula:11e99bbe-5ce3-4a82-9dc7-31ab22e18167}} (steplength){{formula:376e8855-2593-4b59-ad1c-4700161988a1}} Initialize {{formula:30e032c0-db0f-487a-ac4c-2cd5ba60f1a3}} , {{formula:74a1bbd1-bf41-4423-93d1-2be33d82fde9}} ;not converge{{formula:ae1e41b8-6d51-4f82-a64d-a965509a1435}} ; {{formula:d1a0783e-2718-4c31-9f4b-3a3600d465f4}} ;{{formula:3758e3ef-d0de-470a-b71d-9d44646690a7}} ;{{formula:6387d502-336c-4c0e-be99-b5bb38a8044e}} , {{formula:482b58e8-5aa3-46c5-ab3c-a837bb9f26ee}} ;"
] |
[] |
Rules in Play: On the Complexity of Routing Tables and Firewalls
|
A fundamental component of networking infras- tructure is the policy, used inrouting tables and firewalls. Accordingly, there has been extensive study ofpolicies. However, the theory of such policies indicates that the size of thedecision tree for a policy is very large ( O((2n)d), where the policy has nrules and examines d features of packets). If this was indeed the case, theexisting algorithms to detect anomalies, conflicts, and redundancies would notbe tractable for practical policies (say, n = 1000 and d = 10). In this paper,we clear up this apparent paradox. Using the concept of 'rules in play', wecalculate the actual upper bound on the size of the decision tree, anddemonstrate how three other factors - narrow fields, singletons, andall-matches make the problem tractable in practice. We also show how thisconcept may be used to solve an open problem: pruning a policy to the minimumpossible number of rules, without changing its meaning.
|
[
"Owing to the practical importance of policies, there exists aconsiderable body of work devoted to their study. A natural questionis how this paper fits into the context of this research.",
"The first and most obvious area related to our work is the theory ofpolicies. This area includes the design of algorithms and datastructures for fast packet processing (i.e. resolution) as well as forverification of policies. Policies have been represented as tries{{cite:9f61a956579a6fb42a1cc9d269a9f645247b87a2}} and lookup-tables by Waldvogel {{cite:dc75f9b4e7709233400b1fcc79c2a4ecb04a0be2}} and Gupta{{cite:3128906d6be7525993cdfda6db84fcd2bd4d5b8d}}. As preprocessing can be expensive, solutions have beendeveloped by Suri{{cite:d147d109597256d27e63befc14ae4c77bb8cfa72}} using B-trees, and by Sahni andKim{{cite:c5dc610a1cc65e7991ffccd8b31fa923bd910e9c}} using red-black trees and skip lists; these solutionsallow fast update, and also perform (longest prefix) matching in{{formula:81265d97-fa4a-4833-8768-5fb443e8c287}} time.",
"Our work not only advances the theory of policies by providing anexact bound on the size of the policy decision diagram{{cite:84d4990252d3246717bf807f49c2c7f40422632f}}, but also proposes new metrics for the complexity of apolicy, and shows why the algorithms for fast resolution andverification are tractable in practice. Our new metrics (fieldwidth,allprob and oneprob) make it possible to answer if a policy will havea decision diagram of tractable size.",
"More generally, the analysis of policies includes the study ofanomalies {{cite:7aa16bea38b94a8b1b28b2e7a079a17c3839314b}}, inter-rule conflicts {{cite:2f06428c6769a18e20512e35b991f805a3b1374b}},redundancies, and so on. For example, Frantzen {{cite:5add5cad25dc664d7caa97ff6f2c6faaf8ec4b6d}}provides a framework for understanding the vulnerabilities in afirewall, and Blowtorch {{cite:124b0c7e1cc45a43d547b8a57e7a4a265fa54af7}} is a framework to generatepackets for testing. These algorithms depend on the complexity of apolicy, as measured by the size of its decision diagram; our worksheds light on why they are practical to use.",
"Finally, our work impacts the study of high-performance architecturefor fast packet resolution. High-throughput systems, such as backbonerouters, make use of special hardware - ternary content addressablememory{{cite:15c2e8370ebe66db3a7a65619bb791f4855d3213}}, ordinary RAM, pipelining systems {{cite:1e0866c6342e430deb9aa55397a7becef292cb7d}}, andso on. But these systems are very expensive, as well as limited inthe size of policies they can accommodate. Therefore, there is activeresearch on algorithms to optimize policies, such as Liu's TCAMRazor{{cite:15c2e8370ebe66db3a7a65619bb791f4855d3213}}. This paper shows how, using our concept of rules inplay, we can improve upon these algorithms to create truly minimumpolicies (rather than just locally minimal ones)."
] |
[] |
Pixel-wise Segmentation of Street with Neural Networks
|
Pixel-wise street segmentation of photographs taken from a driversperspective is important for self-driving cars and can also support otherobject recognition tasks. A framework called SST was developed to examine theaccuracy and execution time of different neural networks. The best neuralnetwork achieved an $F_1$-score of 89.5% with a simple feedforward neuralnetwork which trained to solve a regression task.
|
[
"Road segmentation is a subproblem of general scene parsing or segmentation. Inscene parsing every object in a scene is classified pixelwise with a label.Whereas in road segmentation often only two classes exist and more assumptionscan be applied.",
"In the first publications, roads were usually annotated by color-basedhistogram approaches and specific model knowledge. Examples are the in 1994introduced approach {{cite:6a5865dc5076b559cad6dee598171fa4a1a6e310}} using the watershed algorithm or{{cite:1eabc5a8ff54ee071928c6eedfe7e60d28141602}} where roads were annotated indirectly by lane markings foundwith a Hough transformation.",
"Later insights of general scene parsing where transferred and more genericapproaches like {{cite:74226c907525511a5b02f94b39f7808121871e1b}} have achieved remarkable results with aMRF and superpixels.",
"The impressive classification results ofCNN like AlexNet {{cite:c44a0d1497180123a8b0f06df1fad4606c73a661}} orGoogLeNet {{cite:6dd7e36de8ad8695d85e34b0069722da6af16e82}} during the Google ImageNet LSVRC-2010contest, made CNN interesting for all kinds of computer vision problemslike segmentation.",
"With {{cite:8b524affcc65727524d55a3e4201b899766db576}} Long and his team introduced a method for generalscene parsing based on FCN and a deconvolutional layer.",
"This approach is used as a blueprint to our implementation, described insec:model. Therefore the main concepts are introduced insec:concept.",
"Instead of creating a new model, they converted existing classificationCNN like AlexNet or GoogLeNet into FCN. The obtained heat maps forevery class were calculated for multiple resolutions and upscaled withdeconvolution layer interpolation to the original resolution. With a fullyconnected convolutional layer in the end, the multiple outputs are combinedinto one classification heat map for every class.",
"In {{cite:51cdc57c70109c3990b16acfea00b450dafaba3a}} a approach is presented, which makes also use of aCNN in combination with deconvolution. In comparison to Long's network,among others it is less deep and uses less convolutional then deconvolutionallayers. Furthermore the input image is divided in multiple patches and for eachpatch a separate neural network was trained. Their model achieved thebest-recorded result on the same data set we use, which is described insec:datasets."
] |
[] |
The height of piecewise-testable languages and the complexity of the
logic of subwords
|
The height of a piecewise-testable language $L$ is the maximum length of thewords needed to define $L$ by excluding and requiring given subwords. Theheight of $L$ is an important descriptive complexity measure that has not yetbeen investigated in a systematic way. This article develops a series of newtechniques for bounding the height of finite languages and of languagesobtained by taking closures by subwords, superwords and related operations. As an application of these results, we show that$\mathsf{FO}^2(A^*,\sqsubseteq)$, the two-variable fragment of the first-orderlogic of sequences with the subword ordering, can only expresspiecewise-testable properties and has elementary complexity.
|
[
"The height of PT languages has been used to measure the difference betweenseparable languages, see e.g. {{cite:582cede1c7e2e49a0dc254fef21d3a0ae29b415f}}.However theliterature is quite poor on the question of estimating the height ofPT languages. Algorithms that decide whether a regular language {{formula:bc7a626a-1ecd-49f1-b6c6-35d5243495af}} (given e.g. by its canonical DFA {{formula:1ca8cb3a-ee7c-4c8b-94bd-c40ec5e80885}} ) is PT usually provide abound on {{formula:0fe555b5-c121-4f4e-a596-135b60a82126}} in terms of {{formula:3da5b4c5-b112-49aa-8671-49e91bb37060}} : recently Klíma andPolák showed that {{formula:c1b73baa-c4da-4a56-8b2f-5bd100bc41f6}} is bounded by the depth of{{formula:ba758e8c-37a2-4e75-9105-c1c0264ff25c}} {{cite:d549f82ece36f7d5f778bf38e1ab8a4f8613a102}}. The currently best bounds on{{formula:d5d013a7-b391-4e5e-8199-094644a6dadc}} based on automata for {{formula:f43581ce-010a-4d5a-a158-bcc50e5cadd3}} have been obtained by Masopust andThomazo {{cite:02b0eedba246e2aacdc94b7ec4cb8a2e6096ba67}}, {{cite:22477a2b1274ec553d477eaf409169631c26a124}}.",
"When {{formula:ace994cc-efe6-4df2-8185-c5b878bef0a1}} is obtained by operations onother languages, very little is known about PT heights. It is clear that{{formula:651d0ff9-eeaf-417b-9a41-497992a8e1ad}} and that {{formula:e7b33dcd-61ad-4f53-acde-823ac0f0304a}} but beyond boolean operations,quotients, and inverse morphisms, thereare very few known ways of obtaining PT languages (see app-closure)."
] |
[] |
An Empirical Study on Sentiment Classification of Chinese Review using
Word Embedding
|
In this article, how word embeddings can be used as features in Chinesesentiment classification is presented. Firstly, a Chinese opinion corpus isbuilt with a million comments from hotel review websites. Then the wordembeddings which represent each comment are used as input in different machinelearning methods for sentiment classification, including SVM, LogisticRegression, Convolutional Neural Network (CNN) and ensemble methods. Thesemethods get better performance compared with N-gram models using Naive Bayes(NB) and Maximum Entropy (ME). Finally, a combination of machine learningmethods is proposed which presents an outstanding performance in precision,recall and F1 score. After selecting the most useful methods to construct thecombinational model and testing over the corpus, the final F1 score is 0.920.
|
[
"According to liu2012survey, the sentiment analysis research mainly started from early 2000 by turney2002thumbs and pang2002thumbs. turney2002thumbs firstly used a few semantically words (e.g., excellent and poor) to label other phrases with the hit counts by queries through search engines. Then, researchers had also proposed several custom techniques specifically for sentiment classification, e.g., the score function based on words in positive and negative reviews {{cite:125b69eb1f1305b258d8938b13b354a5de1b42f3}} and feature weighting schemes used to enhance classification accuracy {{cite:a81ddb79bdd0f5238ee5a5b9b99f7ee07d97a33e}}.Besides, the other situation of sentiment analysis is to represent texts by vectors which indicate these words appear in the text but do not preserve word order. And a machine learning approach will be used for classification in the end. In such way, pang2002thumbs considered classifying documents according to standard machine learning techniques. In addition, subsequent research used more features in learning, making the main task of sentiment classification engineer an effective set of features {{cite:1f3c9d423ae57145a5e8414401050aad511ffbed}}.",
"However, compared to English sentiment analysis, there are relatively few investigations conducted on Chinese sentiment classification until 2005 {{cite:0734b92aa53ce4e1a57e8cbb36410d152102a069}}. li2007experimental presented a study on comparison of different machine learning approaches under different text representation schemes and feature weighting schemes. They found that SVM achieved the best performance. After that, tan2008empirical found 6,000 or bigger for the size of features would be sufficient for Chinese sentiment analysis, and sentiment classifiers were severely dependent on domains or topics.",
"Nowadays, inspired by the availability of large text corpus and the success of deep learning approaches, some researchers (e.g., collobert2011natural, johnson2014effective) deviated from traditional methods and tried to train neural networks such as Convolutional Neural Networks (CNN) for NLP tasks (e.g., named entity recognition and sentiment analysis). Among them, xu2013convolutional and kalchbrenner2014convolutional got some state-of-the-art performance. But the work of collobert2011natural was paid most attention for describing a unified architecture for NLP tasks which learned features by training a deep neural network even when being given very limited prior knowledge. These NLP tasks included part-of-speech tagging, chunking, named-entity recognition, language model learning and semantic role labeling."
] |
[] |
Implementing a Small Parsing Virtual Machine on Embedded Systems
|
PEGs are a formal grammar foundation for describing syntax, and are not hardto generate parsers with a plain recursive decent parsing. However, the largeamount of C-stack consumption in the recursive parsing is not acceptableespecially in resource-restricted embedded systems. Alternatively, we haveattempted the machine virtualization approach to PEG-based parsing. MiniNez,our implemented virtual machine, is presented in this paper with severaldownsizing techniques, including instruction specialization, inline expansionand static flow analysis. As a result, the MiniNez machine achieves both a verysmall footprint and competitive performance to generated C parsers. We havedemonstrated the experimental results by comparing on two major embeddedplatforms: Cortex-A7 and Intel Atom processor.
|
[
"Parsing and syntactic analysis is ubiquitous and an integral part of text processing. Since the development of Lex/Yacc, the parser generator has increasingly adopted as a standard approach to formal grammar engineering. In contexts of data parsing, binpac{{cite:7204ff4a94ffef7fc57092dbf9f00e4f1b913278}} and XMLScreamer{{cite:aee8b8132c59e1b15df454f87d3aab8427c1b01e}} notably extend Yacc to produce efficient protocol parsers and XML parsers, respectively. However, Yacc and its extended tools do not guarantee its availability on resource-restricted embedded systems, while they produce a portable parser code. To our knowledge, the parser generation that is specialized for embedded systems has been unexplored, and then is an important open problem to be addressed.",
"PEGs{{cite:ad91444f8938f30e63d6d03013a81c99b6bfcad5}} are a relatively newer formalism than Context Free Grammars, and have received much popularity among many parser researchers. There are many PEG-based parser tools such as Pappy{{cite:f7374738790fd965791ed0482f548910be1d1174}}, Rats{{cite:3f7807c4e051701e8bc5b4b00d2af39b7cd053b8}}, Mouse{{cite:09341658eeb184e5a37769d040a02b0b06a87010}}, Leg/Peg{{cite:8bdbd7e4f0645ed523337ddfb44f75c2b9fcbbe7}}, PEGjs{{cite:24f003e33e81c25329b018d7af2ac0d3408470b8}}, LPeg{{cite:28e715b03f379b90af43d4e6806c71f1f2169278}}, Waxeye{{cite:dd4da5ffbc3300cc6f3666140391a68e32383472}} and so on. Most of these tools are based on a parser generator as with in Lex/Yacc. An important exception is LPeg, a pattern matching library for Lua{{cite:28e715b03f379b90af43d4e6806c71f1f2169278}}, {{cite:31f26c7eda9823dedb64a3ea69e15971504cc94f}}. As with MiniNez, LPeg has adopted a virtual parsing machine {{cite:6ba644709d648e6ec500e81e87bf2faaf7b6b5a4}} instead of parser generations. MiniNez and LPeg significantly differ in terms of backtracking handling: local static jump vs. global dynamic jump. Since the failure jump address is not pushed on the stack, MiniNez achieves a fewer stack consumption. In addition, we add several specialized downsizing to the MiniNez design and implementation. LPeg, on the other hand, does not work standalone and requires a complementary Lua runtime, resulting in hard comparison on the same condition."
] |
[] |
Two Hops or More: On Hop-Limited Search in Opportunistic Networks
|
While there is a drastic shift from host-centric networking tocontent-centric networking, how to locate and retrieve the relevant contentefficiently, especially in a mobile network, is still an open question. Mobiledevices host increasing volume of data which could be shared with the nearbynodes in a multi-hop fashion. However, searching for content in thisresource-restricted setting is not trivial due to the lack of a content index,as well as, desire for keeping the search cost low. In this paper, we analyze alightweight search scheme, hop-limited search, that forwards the searchmessages only till a maximum number of hops, and requires no prior knowledgeabout the network. We highlight the effect of the hop limit on both searchperformance (i.e., success ratio and delay) and associated cost along with theinterplay between content availability, tolerated waiting time, networkdensity, and mobility. Our analysis, using the real mobility traces, as well assynthetic models, shows that the most substantial benefit is achieved at thefirst few hops and that after several hops the extra gain diminishes as afunction of content availability and tolerated delay. We also observe that thereturn path taken by a response is on average longer than the forward path ofthe query and that the search cost increases only marginally after several hopsdue to the small network diameter.
|
[
"While the DTN literature has many proposals for message dissemination, which exploit the information about the network such as (estimated)pairwise node contactrates, node centrality, communities, and social ties, efficient mechanisms for content search remain largely unexplored.In a sense, this is reasonable as search can be considered as a two-step message delivery: query routing on the forward path and the response routing on the return path. The forward path is less certain as the content providers are unknown. In this regard, the return path is less challenging as the target node (and a recent path to reach it) is already known. However, routing the response looks for a particular node, whereas the forward search is for a subset of nodes whose cardinality is proportional to the content availability.Thus, search requires special treatment rather than being an extension of the message dissemination.",
"Two questions for an efficient search are (i) which nodes may have the content {{cite:7a5360394442fcddf3213521bd52f515f0bc96ce}} and (ii) when to stop a search {{cite:08e8e835b6b3a8f61f62159f24cf64b6187bbca4}}. The former question requires assessment of each node in terms of its potential of being a provider for the sought content. For example, seeker assisted search (SAS) {{cite:7a5360394442fcddf3213521bd52f515f0bc96ce}} estimates the nodes in the same community to have higher likelihood of holding the content as people in the same community might have already retrieved the content.Given that people sharing a common interest come together at a certain “space\" (e.g., office, gyms), {{cite:60c0a74d614c7265ab13c566ce2f58c864be57e6}} defines geo-community concept and matches each query with a particular geo-community.Hence, the first question boils down to selecting relays with high probability of visiting the target geo-community.As {{cite:60c0a74d614c7265ab13c566ce2f58c864be57e6}} aims to keep the search cost minimal,searching node employs two-hop routing and determines the relays, which thereby reduces the issue of when to stop the search.Deciding when to stop search is nontrivial as the search follows several paths and whether the searching node has already discovered a response is not known by the relaying nodes. Hyytiä et al. {{cite:2aa53f824893201600d577363042e55e9459e2a2}} model the expected search utility with increasing hop count and then finds the optimal hop, while {{cite:08e8e835b6b3a8f61f62159f24cf64b6187bbca4}} estimates the number of nodes having received a query and possible responses by using the node degrees.In our work, we showed that simple schemes via information sharing can stop search timely and maintain similar performance to that of an oracle due to the small world nature of the studied networks.",
"Different from all these above works, main focus of our paper is more on a fundamental question:how does the hop limitation affect the search performance?While this question has beenexplored in general networking context,content-centricity and the time constraintsrequire a better understanding of flooding in the context of search.Therefore, we first provided insights on search on a simplified setting and nextanalyzed the effect of various parameters, e.g., time and real mobility traces, via extensive simulations.",
"In the literature, several works focus on two-hop forwarding in which the source node replicates the message to any relay and the relays can deliver the message only to the final destination {{cite:df411250ae131ea34dd09b310deaea9d67d71517}}, {{cite:24fcfa531a235188f0c7045a4f85f80e4789f11d}}, {{cite:a7cf3e72a5b7d1ad67ff6bc7c0f7c88f32610601}}.Grossglauser et al. in their seminal work {{cite:df411250ae131ea34dd09b310deaea9d67d71517}} show that two hops are sufficient to achieve the maximum throughput capacity in an ideal network with nodes moving randomly.The capacity increase is facilitated by the reduced interference on the links from source to relay and relay to the destination.Chaintreau et al. {{cite:24fcfa531a235188f0c7045a4f85f80e4789f11d}} assess this two-hop forwarding scheme in a DTN scenario with power-law inter-contact times andemploy an oblivious forwarding algorithm (e.g., memoryless routers that do not use any context information such as contact history).Similarly, {{cite:d7bcb2bed84541b659032e4e62282cca62d7e692}} focuses on a DTN with power-law distributed inter-contact times and derives the conditions (i.e. range of Pareto shape parameter) under which message delivery has a finite delay bound for both two-hop and multi-hop oblivious algorithms.The authors show that“as long as the convergence of message delivery delay is the only concern, paths longer than two-hops do not help convergence” as two hops are sufficient to explore the relaying diversity of the network {{cite:d7bcb2bed84541b659032e4e62282cca62d7e692}}.Another work supporting two-hop schemes is {{cite:60c0a74d614c7265ab13c566ce2f58c864be57e6}} which shows that two-hop search is favourable for opportunistic networks –a resource-scarce setting,as “one-hop neighbors are able to cover the most of the network in a reasonable time\" in a network with sufficiently many mobile nodes.In our work, we include those cases when longer paths still yield (some) performance improvement.Finally, our work supports the conclusion of {{cite:0bd0154356b689e19468b57a8b6e68706bd4bedc}}, which theoretically proves the small-world in human mobile networks.",
"Our work is closely related to the {{formula:b0353fcc-3715-4934-8a28-92b1409c6b54}} -hop flooding{{cite:f8c34d54e4014bc92f7de1eb20c7b204642881c6}} which models the spread of flooded messages in a random graph.Unlike {{cite:f8c34d54e4014bc92f7de1eb20c7b204642881c6}}, we focus on content search in a realistic setting, and using real mobility traces (both a human contact network and a vehicular network) we provide insights on the effect of increasing hop count on the search success, delay, and cost under various content availability and tolerated waiting time settings.Despite the differences, it is worthwhile noting that our results agree with the basic conclusions of {{cite:f8c34d54e4014bc92f7de1eb20c7b204642881c6}}."
] |
[] |
Diversity Networks: Neural Network Compression Using Determinantal Point
Processes
|
We introduce Divnet, a flexible technique for learning networks with diverseneurons. Divnet models neuronal diversity by placing a Determinantal PointProcess (DPP) over neurons in a given layer. It uses this DPP to select asubset of diverse neurons and subsequently fuses the redundant neurons into theselected ones. Compared with previous approaches, Divnet offers a moreprincipled, flexible technique for capturing neuronal diversity and thusimplicitly enforcing regularization. This enables effective auto-tuning ofnetwork architecture and leads to smaller network sizes without hurtingperformance. Moreover, through its focus on diversity and neuron fusing, Divnetremains compatible with other procedures that seek to reduce memory footprintsof networks. We present experimental results to corroborate our claims: forpruning neural networks, Divnet is seen to be notably superior to competingapproaches.
|
[
"Due to their large number of parameters, deep neural networks typically have a heavy memory footprint. Moreover, in many deep neural network models parameters show a significant amount of redundancy {{cite:8f051910157d9e530321b5ef7b0b9acc985af09e}}. Consequently, there has been significant interest in developing techniques for reducing a network's size without penalizing its performance.",
"A common approach to reducing the number of parameters is to remove connections between layers. In {{cite:f2f18eb315af15b98fa701b20440252527056f93}}, {{cite:c896f790afe6e5349a3b1cd37b29e4be18adc675}}, connections are deleted using information drawn from the Hessian of the network's error function. {{cite:0fb2a9d93cf99c75966e65bbb712d9963041e5ba}} reduce the number of parameters by analyzing the weight matrices, and applying low-rank factorization to the final weight layer. {{cite:fb006cc94901fdcf4155e9f9ebfd8412e8b446de}} remove connections with weights smaller than a given threshold before retraining the network. These methods focus on deleting parameters whose removal influences the network the least, while Divnet seeks diversity and merges similar neurons; these methods can thus be used in conjunction with ours.Although methods such as {{cite:f2f18eb315af15b98fa701b20440252527056f93}} that remove connections between layers may also delete neurons from the network by removing all of their outgoing or incoming connections, it is likely that the overall impact on the size of the network will be lesser than approaches such as Divnet that remove entire neurons: indeed, removing a neuron decreases the number of rows or columns of the weight matrices connecting the neuron's layer to both the previous and following layers.",
"Convolutional Neural Networks {{cite:762a7d6a3f2580c6aacf721fab5a89af050e7d2e}} replace fully-connected layers with convolution and subsampling layers, which significantly decreases the number of parameters. However, as CNNs still maintain fully-connected layers, they also benefit from Divnet.",
"Closer to our approach of reducing the number of hidden neurons is {{cite:2b41cba634571ce421a5f71a9351577fea510a8c}}, which evaluates each hidden neuron's importance and deletes neurons with the smaller importance scores. In {{cite:8116a86299a91c663a3bb723b05e3d4b58af2843}}, a neuron is pruned when its weights are similar to those of another neuron in its layer, leading to a weight update procedure that is somewhat similar in idea (albeit simpler) to our reweighting step: where {{cite:8116a86299a91c663a3bb723b05e3d4b58af2843}} removes neurons with equal or similar weights, we consider the more complicated task of merging neurons that, as a group, perform redundant calculations based on their activations.",
"Other recent approaches consider network compression without pruning: in {{cite:4fbd4e394815521bc5ebb98a7006d73809a5388b}}, a new, smaller network is trained on the outputs of the large network; {{cite:60b895f37bb9713832b25d8e546d68ac32da94b3}} use hashing to reduce the size of the weight matrices by forcing all connections within the same hash bucket to have the same weight. {{cite:9ef98b057a5b6888638154fa5937b8cfc7eac791}} and {{cite:cb2d9c1bc091748b3206ef9f588ee2a52be1e15d}} show that networks can be trained and run using limited precision values to store the network parameters, thus reducing the overall memory footprint.",
"We emphasize that Divnet's focus on neuronal diversity is orthogonal and complementary to prior network compression techniques. Consequently, Divnet can be combined, in most cases trivially, with previous approaches to memory footprint reduction."
] |
[] |
Recurrent Neural Networks Hardware Implementation on FPGA
|
Recurrent Neural Networks (RNNs) have the ability to retain memory and learndata sequences. Due to the recurrent nature of RNNs, it is sometimes hard toparallelize all its computations on conventional hardware. CPUs do notcurrently offer large parallelism, while GPUs offer limited parallelism due tosequential components of RNN models. In this paper we present a hardwareimplementation of Long-Short Term Memory (LSTM) recurrent network on theprogrammable logic Zynq 7020 FPGA from Xilinx. We implemented a RNN with $2$layers and $128$ hidden units in hardware and it has been tested using acharacter level language model. The implementation is more than $21\times$faster than the ARM CPU embedded on the Zynq 7020 FPGA. This work canpotentially evolve to a RNN co-processor for future mobile devices.
|
[
"Co-processors for accelerating computer vision algorithms and CNNs have been implemented on FPGAs. A system that can perform recognition on mega-pixel images in real-time is presented in {{cite:0b090d15db38495641010cdb1da38ea6061aa991}}. A similar architecture for general purpose vision algorithms called neuFlow is described in {{cite:2dd619f801f962bddd013c33c356a987b93bb373}}. neuFlow is a scalable architecture composed by a grid of operation modules connected with an optimized data streaming network. This system can achieve speedups up to {{formula:6214d984-675f-47fa-92b3-3845808c275d}} in end-to-end applications.",
"An accelerator called nn-X for deep neural networks is described in {{cite:2957ba9b0c2b411de6a3cfef6682417d8762ca30}}, {{cite:3aa6b05644b29b7da95c67299477de485faea870}}, {{cite:e8e6b40ed2833a00587fd4cff4ca9af59788b853}}, {{cite:3f92fb6093227b0e13d18904cbb90166d473147b}}. nn-X is a high performance co-processor implemented on FPGA. The design is based on computational elements called collections that are capable of performing convolution, non-linear functions and pooling. The accelerator efficiently pipelines the collections achieving up to 240 G-op/s.",
"RNNs are different from CNNs in the context that they require a different arrangement of computation modules. This allows different hardware optimization strategies that should be exploited. A LSTM learning algorithm using Simultaneous Perturbation Stochastic Approximation (SPSA) for hardware friendly implementation was described in {{cite:0d4ae5669c75f665bd9dd192e88823df1777a959}}. The paper focuses on transformation of the learning phase of LSTM for FPGA.",
"Another FPGA implementation that focus on standard RNN is described by {{cite:acda898fcc3f3e9fc4e9501f33e2b44098b3fa1a}}. Their approach was to unfold the RNN model into a fixed number of timesteps {{formula:d498ddae-4b41-4106-9aec-f16a92aa9374}} and compute them in parallel. The hardware architecture is composed of a hidden layer module and duplicated output layer modules. First, the hidden layer serially processes the input {{formula:e289b656-6477-476c-a8b8-15d25f243739}} for {{formula:77d0f909-cf82-48d6-aedf-7fa90756a179}} timesteps. Then, with the results of the hidden layer, the duplicated logic computes output {{formula:ec2bad2e-4e51-4233-8a30-d1d7aecc7ff2}} for {{formula:8280bfc8-f66f-49b8-9c1d-8ebb563bfffc}} timesteps in parallel.",
"This work presents a different approach of implementing RNN in FPGA, focusing the LSTM architecture. It is different than {{cite:acda898fcc3f3e9fc4e9501f33e2b44098b3fa1a}}, in the sense that it uses a single module that consumes input {{formula:573224af-6068-41e4-ada1-718819b5453c}} and previous output {{formula:d0027ad2-c2a5-43db-8f11-c523e662fd81}} simultaneously."
] |
[] |
Empirical Research and Automatic Processing Method of Precision-specific
Operation
|
Significant inaccuracy often occurs during the process of mathematicalcalculation due to the digit limitation of floating point, which may lead tocatastrophic loss. Normally, people believe that adjustment of floating-pointprecision is an effective way to solve this problem, since high-precisionfloating-point has more digits to store information. Thus, it is a prevalentmethod to reduce the inaccuracy in much floating-point related research, thatperforming all the operations with higher precision. However, we discover thatsome operations may lead to larger error in higher precision. In this paper, wedefine this kind of operation that generates large error due to precisionadjustment a precision-specific operation. Furthermore, we propose alight-weight searching algorithm for detecting precision-specific operationsand figure out an automatic processing method to fixing them. In addition, weconducted an experiment on the scientific mathematical library of GLIBC. Theresult shows that there are many precision-specific operations, and our fixingapproach can significantly reduce the inaccuracy.
|
[
"To resolve the aforementioned problem,recent research works employ the techniques of precision adjusting.",
"Increasing precision is a common approach to ensure the calculationresults more close to the “true” real number in the scientific computation.For instance, climate modeling requires complicated mathematical computation and simulation.Priori work {{cite:dee228927d421b751c181ca3672d81d66614e64b}} employed a double-double arithmetic for a climate modeling code,which reduce the numerical variability of the entire system dramatically.By using higher precision, they significantly improve numerical reproducibilityand stability in Parallel Applications.Another research is in the supernova simulation area.P. H. Hauschildt and E. Baron {{cite:4ddb62cd05949aa4da4174a94282de569522d467}} used double-double (128-bit or 31-digit)and quad-double (256-bit or 62-digit) types solve the expanding stellar atmosphere problem successfully.Extending the length of digits to store a more exact number, in another word,processing the program in higher precision, is a popular approach to decrease the potential error.",
"Sometimes, speed is more important than accuracy to an application.In this case, researchers prefer to decreasing precision to perform more operations in one second. {{cite:7a028c497797f895a9c3d684347e5ac13faa0a90}} {{cite:ddb8ae4d203995ac776597b10a8a675fa487151e}}For instance, {{cite:d6156c270ecdb4d585f1dc12c5c61344b5b6c481}} proposed a framework that performs better by decreasing precision.They designed a breadth first search algorithm to automate identification of code regionsthat can use lower precision.After that, they used binary instrumentation and modification to build mixed-precision configurationand get some transformations of the original program.Then, they kept decreasing the precision of different transformations until the result is qualified enough to pass the efficiency test.Their work limits the accuracy loss of decreasing precision."
] |
[] |
Towards Arbitrary-View Face Alignment by Recommendation Trees
|
Learning to simultaneously handle face alignment of arbitrary views, e.g.frontal and profile views, appears to be more challenging than we thought. Thedifficulties lay in i) accommodating the complex appearance-shape relationsexhibited in different views, and ii) encompassing the varying landmark pointsets due to self-occlusion and different landmark protocols. Most existingstudies approach this problem via training multiple viewpoint-specific models,and conduct head pose estimation for model selection. This solution isintuitive but the performance is highly susceptible to inaccurate head poseestimation. In this study, we address this shortcoming through learning anEnsemble of Model Recommendation Trees (EMRT), which is capable of selectingoptimal model configuration without prior head pose estimation. The unifiedframework seamlessly handles different viewpoints and landmark protocols, andit is trained by optimising directly on landmark locations, thus yieldingsuperior results on arbitrary-view face alignment. This is the first study thatperforms face alignment on the full AFLWdataset with faces of different viewsincluding profile view. State-of-the-art performances are also reported onMultiPIE and AFW datasets containing both frontaland profile-view faces.
|
[
"Many approaches have been proposed for solving the face alignment problem.The classic methods include the Active Appearance Model (AAM) {{cite:78498c0d31a597250d46b10acec9a0994c695c48}} and its variants.AAM models faces by both appearance and shape information and optimises them in a holistic manner.Facial parts detection based method {{cite:adba6321c3a0ed77d0b55f9ad9d36b88edcdcd2f}}, also known as Constrained Local Model, aim to detect facial parts with several part detectors, and use global constraints to estimate reasonable shapes.Discriminative regression methods {{cite:b561424b3b3f54bd363d226bde93bfe483134750}}, {{cite:bc5079e7a7ed1a12251ae23b7ec71262a0a04999}}, {{cite:6630cde7df76d3e34423565ea865a36fc6d5064f}}, {{cite:e5ada708455f8e1d30024dcb3f2106e37ffd2945}} regress expected shape update based on local appearance information extracted from all current estimated landmarks.Such methods have achieved state-of-the-art accuracy.Our framework allows these state-of-the-art approaches to be applicable to non-frontal cases via recommendation over multiple models.A number of existing studies propose to solve the multi-view face alignment problem {{cite:7991e4e0465b34a5b072a85ffc84ea92477366b6}}, {{cite:f0fd734d0e3fd56b22a91ed4879cab8b4ec23ec5}}. These studies differ significantly from our work since they confine the yaw angle of faces within the range of {{formula:50b066c8-68a8-47ce-b2da-52155286dd7a}} 45{{formula:115dc63a-0eb8-49d5-ac46-cda30713dc57}} {{cite:7991e4e0465b34a5b072a85ffc84ea92477366b6}}.In {{cite:f0fd734d0e3fd56b22a91ed4879cab8b4ec23ec5}}, the dataset is chosen from the 300-W dataset {{cite:f2d0b97b1182e5588bfa4123b147a644be89dd7b}} with an approximate range of {{formula:3e972ed3-adbb-4a0b-9038-f646cf3e123e}} 45{{formula:a8feb78b-65c3-453d-b317-6d0a38b39a3c}} , which is far smaller than the head pose range considered in this study.",
"Our work is closer to a few alignment studies that address the face alignment problem in a multimodal manner:",
"Conditional Regression Forest {{cite:2d8b5b76be5f36174509c777f3107206ec696fc3}}. In this work, the facial points are estimated after a rough estimation of head pose.The head pose is estimated according to the global appearance of the face image and regarded as the prior probability of each alignment model.This method works quite well and gains improvement on traditional face alignment benchmarks, LFPW {{cite:adba6321c3a0ed77d0b55f9ad9d36b88edcdcd2f}} or LFW {{cite:2d8b5b76be5f36174509c777f3107206ec696fc3}}.When this method is extended to dataset with extremely large pose variations, AFLW, the method fails due to erroneous head pose estimation, no matter how well each face alignment model is trained.",
"Iterative Multi-Output Random Forests {{cite:d8ef126083f79560affa3e02362abed0a50b339b}}. This study extends {{cite:2d8b5b76be5f36174509c777f3107206ec696fc3}} by iteratively re-estimating the head pose and facial expression based on the current estimated shape, and then re-estimating the facial landmarks with the updated prior model probability.Further improvement is achieved in this work compared to {{cite:2d8b5b76be5f36174509c777f3107206ec696fc3}}.Similar to {{cite:2d8b5b76be5f36174509c777f3107206ec696fc3}}, the algorithm begins with head pose estimation based on global appearance.When applied to faces with large pose variation, if the initial head pose estimation is incorrect, the error is rather hard to be rectified during iterative optimisation since the initial shape is estimated with the wrong prior probability.",
"Consensus of Regressors {{cite:dcf9500fe1f4db07bb60fa64d5694408237bb778}}. The study applies a mixture of models to handle different types of occluded faces.The method works well on occlusion benchmarks like COFW {{cite:3b7b6f3e80ffe660b5b0efd93718a56440d5e898}}.However, the method yields poor performance on profile-view cases because all regression models assume a limited pose variation.From our experiments, we observe that even if models are provided to handle large pose variations, suboptimal results may be still generated due to the failure in drawing consensus amongst recommended landmarks with large spatial variations.",
"Model Recommendation. Model recommendation has been proposed for action recognition {{cite:31a85f0f35b094378d76691d7ce5323f79c145bc}}. The study shows that a recommendation system is important for choosing one model out of a large set of possibilities. Their method follows the conventional collaborative filtering scheme, where user's rating is required on a small subset of models, and the goal is to use that subset of ratings to predict the ratings of remaining models. This is in contrast to our task where user rating is not available."
] |
[] |
Spatio-temporal video autoencoder with differentiable memory
|
We describe a new spatio-temporal video autoencoder, based on a classicspatial image autoencoder and a novel nested temporal autoencoder. The temporalencoder is represented by a differentiable visual memory composed ofconvolutional long short-term memory (LSTM) cells that integrate changes overtime. Here we target motion changes and use as temporal decoder a robustoptical flow prediction module together with an image sampler serving asbuilt-in feedback loop. The architecture is end-to-end differentiable. At eachtime step, the system receives as input a video frame, predicts the opticalflow based on the current observation and the LSTM memory state as a densetransformation map, and applies it to the current frame to generate the nextframe. By minimising the reconstruction error between the predicted next frameand the corresponding ground truth next frame, we train the whole system toextract features useful for motion estimation without any supervision effort.We present one direct application of the proposed framework inweakly-supervised semantic segmentation of videos through label propagationusing optical flow.
|
[
"Architectures based on LSTM cells {{cite:b1d649b92b70c3fe9d24f2fc99c2ff2bed45eb29}} have been very successful in various tasks involving one-dimensional temporal sequences: speech recognition {{cite:717b34ff2eac9011f2a32668310b9d8faa27cda7}}, machine translation {{cite:518a2ae3689339c40d390557143e1d87bd237b46}}, music composition {{cite:f63b71f14b64a578778b44f82bc4950ae0631261}}, due to their ability to preserve information over long periods of time. Multi-dimensional LSTM networks have been proposed to deal with (2D) images {{cite:fa9d7afdbb3eebf55c29a49cb2a7c567cbd8c54d}} or (3D) volumetric data {{cite:1b21feacc426b6e6522ac50fc6b33e8583ad9865}}, treating the data as spatial sequences. Since in our work we aim at building a visual short-term memory, customised LSTM cells that deal with temporal sequences of spatial data represent a natural choice.",
"Recently, {{cite:de7c3ecbe814e3e554e1bccc871d738c319b5b9a}} proposed an LSTM-based video autoencoder, which aims at generating past and future frames in a sequence, in an unsupervised manner. However, their LSTM implementation treats the input as sequences of vectors by flattening the frames or by using 1D frame representations produced after the last fully-connected layer of a convolutional neural network. This results in a large number of parameters, since the activations of the LSTM cells use fully-connected linear operations, which are not necessarily useful for 2D images, since natural image statistics indicate only local correlations. Another difference is in the architecture setup. We are not interested in training a black-box to produce past and future frames in the sequence. Instead we aim at a transparent setup, and train a more generic memory module together with specialised modules able to decode the memory content and on which appropriate constraints (sparsity, smoothness) can be imposed.",
"Our approach is partially related to optical flow estimation works like DeepFlow {{cite:14c2f0da41a3668f3b8cc1d6166e91c9d192a878}} and FlowNet {{cite:1238de9d26c5ec50ad0894cfbe5778bf828d1581}}. However, these methods use purely supervised training to establish matches between consecutive pairs of frames in a sequence, and then apply a variational smoothing. Our architecture is end-to-end differentiable and integrates a smoothness penalty to ensure that nearby pixels follow a locally smooth motion, and requires no labelled training data.",
"As mentioned in the previous section, our work is similar in spirit to {{cite:c9933813a3b5918922f393ce652753d03ac0d18d}}, by establishing a direct link between vision and motion, in an attempt to reduce supervision effort for high-level scene understanding tasks. However, instead of relying on data provided by an inertial measurement unit from which we can estimate only a global transformation representing the ego-motion, we predict a dense flow map by integrating visual stimuli over time using a memory module, and then we use a built-in feedback loop to assess the prediction. The dense flow map is useful to explain dynamic environments, where different objects can follow different motion models, hence a global transformation is not sufficient {{cite:7332f1ce6f3399d8f422b81cb87a06de0e4bfa95}}."
] |
[] |
Good, Better, Best: Choosing Word Embedding Context
|
We propose two methods of learning vector representations of words andphrases that each combine sentence context with structural features extractedfrom dependency trees. Using several variations of neural network classifier,we show that these combined methods lead to improved performance when used asinput features for supervised term-matching.
|
[
"Many different approaches to learning vector representations of words have been taken. Methods trained from large text examples by relying directly or indirectly on word co-occurrence counts date back to latent semantic analysis {{cite:7eada00964648a3ae775790a3e07f287981184dc}}. Recent examples include the window-based approach of Mikolov et al. (2013c) and the GloVe method {{cite:4e23231ff8f576a2848becb7c204fa12fb26717d}}. There has also been work incorporating prior knowledge with bag-of-words context {{cite:c636dc468119bf1053ac8788aa4fc504a7b2d6aa}}.",
"It is also possible to perform unsupervised learning on structured information about the underlying entities or concepts being represented, such as the relationships encoded in an online knowledge base. Approaches for utilizing this type of training data include the use of neural networks and/or kernel density estimation {{cite:d49cbc7f474885720fc75a14874a6aefc537d96b}} and encoding relations as translations on hyperplanes {{cite:8a688ffa9cdcba18ed7d51598c23bc2b751a7410}}.",
"The most direct precursor to our work is the well-known Word2Vec algorithm {{cite:f5c57e76aa9daa8e56eab8cbfe315ad5daa44e2f}}, {{cite:d15e7c4598b0bb081b9a92cf84c97fbe8899f8bb}}, {{cite:c704f14861bd6264fa1cb71558ef57d886714661}}. It consists of a simple single-layer neural model where the goal is to maximize the ability of the word representations to predict neighbors of the word in a large raw-text corpus. It was observed by Levy and Goldberg (2014) that this algorithm could be generalized to use any type of discrete features with which words could be paired as the “contexts” to be predicted."
] |
[] |
SceneNet: Understanding Real World Indoor Scenes With Synthetic Data
|
Scene understanding is a prerequisite to many high level tasks for anyautomated intelligent machine operating in real world environments. Recentattempts with supervised learning have shown promise in this direction but alsohighlighted the need for enormous quantity of supervised data --- performanceincreases in proportion to the amount of data used. However, this quicklybecomes prohibitive when considering the manual labour needed to collect suchdata. In this work, we focus our attention on depth based semantic per-pixellabelling as a scene understanding problem and show the potential of computergraphics to generate virtually unlimited labelled data from synthetic 3Dscenes. By carefully synthesizing training data with appropriate noise modelswe show comparable performance to state-of-the-art RGBD systems on NYUv2dataset despite using only depth data as input and set a benchmark ondepth-based segmentation on SUN RGB-D dataset. Additionally, we offer a routeto generating synthesized frame or video data, and understanding of differentfactors influencing performance gains.
|
[
"Prior work on per-pixel indoor image labelling has been due to NYUv2 and SUN RGB-D datasets. The work of {{cite:c8354b54273fe2e36db443e148ccb04c0525d153}} was one of the first in the direction and built on top of a deep learning framework trained on NYUv2. However, it achieved only modest performance on the test data. Subsequently, {{cite:7a49f4f36f6d24a1cb3a17e3649f5ec3f1f056b7}} and {{cite:e9598dafbae7967c615e3f95caf66b1d1a95ef82}} have improved the performance, again with deep learning inspired methods. We think that the potential of these new methods is yet to be explored fully and that the lack of training data is the primary hindrance. Both NYUv2 and SUN RGB-D are limited in their sizes and only provide per-pixel labels for low quality raw depth-maps and corresponding RGB frames — missing data and noise in raw sensor measurements exacerbate the problem even more. Also, since most of the labelling relies on human labour, missing labels and mislabelled data are very common, as shown in Fig. REF . This is inevitable as labelling for humans can be a tiring process and sometimes comes with a considerable monetary cost.",
"Ideally, one would like to have a fully labelled 3D model for every scene to generate annotations from arbitrary viewpoints but this is clearly missing in both the datasets. SUN3D {{cite:fca5a1ac3be2f86c9fbc4f1dc98e3b57a500a17a}} goes in the direction of providing annotated video sequences together with 3D point clouds obtained with SfM. However, they only provide 8 such sequences. The NYUv2 dataset provides a large number of videos, but only provide one annotated frame per video. Additionally, the videos are captured without 3D reconstruction in mind and therefore not suitable for generating accurate 3D reconstructions or annotations, as observed from our own experiments. Furthermore, fusion of raw depth images within a temporal window can provide smooth depth-map to aid the segmentation as opposed to noisy raw depth-maps. {{cite:fca5a1ac3be2f86c9fbc4f1dc98e3b57a500a17a}} fuse the raw depth maps but again they are limited by the overall variety of annotated sequences. On the other hand, NYUv2 and SUNRGBD do not provide fused depth measurements. Fortunately, both datasets provide an inpainted version of the raw depth maps which has been used in {{cite:c8354b54273fe2e36db443e148ccb04c0525d153}}, {{cite:7a49f4f36f6d24a1cb3a17e3649f5ec3f1f056b7}}, {{cite:e9598dafbae7967c615e3f95caf66b1d1a95ef82}} on per-pixel scene understanding. In this work, we provide a new way of generating potentially unlimited labelled training data with perfect ground truth inspired from computer graphics. We build our own new library of synthetic indoor scenes, called SceneNet and generate data to train our deep learning algorithm for per-pixel labelling.{{figure:b54643a0-92d7-4ee1-98dd-f2cfea77b72c}}",
"The idea of using synthetic scenes has existed in the past, in particular, {{cite:926eaea645cb866f6f8ccc8ea9c079a4cdca31a0}}, who released a number of scenes targetted towards the application of scene retrieval, which could potentially be used in the problem we are interested in. However, those scenes are small scale of the order of 4m{{formula:2ceb92e4-9bec-4355-bb5c-48c429505dbd}} 3m{{formula:db29b2db-c0b9-4db9-b296-c01589e45058}} 3m, and contain only one or two instances of characteristic objects that define the scene e.g. only one desk and monitor in the office room. On the other hand, object repositories have existed for a long time now particularly the famous Google Warehouse {{cite:517901776069c3c6f1810785ab09168076b46663}} and now ModelNet {{cite:c999d01489fcf26fc85e2b4649d42567bafe2346}} and ShapeNet {{cite:f6ca5f675b7b1f9a5d566067cd99253b4d75eac4}}. Unfortunately, object repositories are not directly useful for the problem we are targetting. Therefore, we compile a number of basis scenes downloaded from the internet and synthesize virtually unlimited number of new scenes by placing objects sampled from these object repositories with various constraints on their joint placement, using simulated annealing {{cite:6b6f062e539aa4a298dc6981abc05f424284be2f}}. In doing so, we not only add variety to our dataset we also obtain free annotations automatically thus completely removing human labellers (or MTurk) from the loop.{{figure:d692ddd5-d512-459d-b8e6-9471f19941fb}}"
] |
[] |
Industrial Experiences with a Formal DSL Semantics to Check Correctness
of DSL Transformations
|
A domain specific language (DSL) abstracts from implementation details and isaligned with the way domain experts reason about a software component. Thedevelopment of DSLs is usually centered around a grammar and transformationsthat generate implementation code or analysis models. The semantics of thelanguage is often defined implicitly and in terms of a transformation toimplementation code. In the presence of multiple transformations from the DSL,the consistency of the generated artifacts with respect to the semantics of theDSL is a relevant issue. We show that a formal semantics is essential forchecking the consistency between the generated artifacts. We exploit the formalsemantics in an industrial project and use formal techniques based onequivalence checking and model-based testing for consistency checking. Wereport about our experience with this approach in an industrial developmentproject.
|
[
"In {{cite:450b6a87e791e3922828f668014398b305bf5765}} the authors prototype the semantics of a DSL called SLCO. The core semantics is described in terms of a transformation to an intermediate language called CS. Afterwards, CS models are transformed to labeled transition systems and are inspected manually or analyzed by existing tools. During the development of SLCO, the authors have implemented a number of transformations, including transformations from SLCO to SLCO models with equivalent behaviors.The authors suggest that the prototype semantics can be used to compare the underlying labeled transition systems of the source and target models.",
"The B method {{cite:40dc13e714a1eccd31ada4a021a5c4ab6af0b0c6}} has been used in {{cite:45e1381a0ec3ab74a351a3e994090521a065c1af}} to develop process schedulers based on specifications in a DSL. The information given by a DSL model is taken into account at several refinement steps in B machines. The authors also introduce a decidable logic for expressing proof obligations of the refinement steps. This allows them to automatically prove the refinements.",
"The mentioned studies focus on validating refinement steps. In contrast, we offer various formal techniques to assess the semantic correctness of different types of artifacts.",
"In {{cite:cbada12f9920f4877b66662f2330a18201f34c7d}} a combination of model-based techniques is used to develop a software bus in a two-phase process. In the first phase, an mCRL2 model of the component is created and validated through simulation. After developing the component, the mCRL2 model is used for model-based testing of the implementation. In the second phase, different properties are verified against the mCRL2 model. The model is improved based on the results and is used for model-based testing against a second implementation. In comparison, our approach is centered around domain-specific models and artifacts generated from them.",
"In {{cite:cf05bced98d5cd68ccf8298fd1179577eec26475}} the POOSL language is used to create a requirements model and a preliminary design model for Pedal Handling. These models are translated to CSP {{cite:0869996262fb011c41f36b38f79bba0146311a0b}} and compared using the refinement checker FDR2 {{cite:a6c4ac951ff47eb89a028043ad3daef1d7619743}}. In {{cite:cf05bced98d5cd68ccf8298fd1179577eec26475}} it is assumed that Pedal Handling consists of an input buffer with certain scheduling rules and a component that processes the inputs from the buffer to make decisions about X-ray requests to the tubes. Hence, the scope of the models is wider compared to the work presented in this paper. As a result, refinement checking is not always feasible. In contrast, specifications in the Pedal Handling DSL focus on the decision making aspect of the component. This reduces the complexity of the models and makes it feasible to compare behaviors by equivalence checking.",
"In Section , we have used model-based testing to validate an implementation of a DSL model. As an alternative, one can use a model learning approach. Model learning techniques extract an automaton model for an implementation by systematically performing tests on it and observing its behavior {{cite:6d630bfbc05ff1e57b00a6257d08679c5fe8b56e}}. The extracted model can be compared with the labeled transition system described by a DSL model."
] |
[] |
Dynamic Parallel and Distributed Graph Cuts
|
Graph-cuts are widely used in computer vision. In order to speed up theoptimization process and improve the scalability for large graphs, Strandmarkand Kahl introduced a splitting method to split a graph into multiple subgraphsfor parallel computation in both shared and distributed memory models. However,this parallel algorithm (parallel BK-algorithm) does not have a polynomialbound on the number of iterations and is found non-convergent in some cases dueto the possible multiple optimal solutions of its sub-problems. To remedy this non-convergence problem, in this work we first introduce amerging method capable of merging any number of those adjacent sub-graphs whichcould hardly reach an agreement on their overlapped region in the parallel BKalgorithm. Based on the pseudo-boolean representations of graph cuts,ourmerging method is shown able to effectively reuse all the computed flows inthese sub-graphs. Through both the splitting and merging, we further propose adynamic parallel and distributed graph-cuts algorithm with guaranteedconvergence to the globally optimal solutions within a predefined number ofiterations. In essence, this work provides a general framework to allow moresophisticated splitting and merging strategies to be employed to further boostperformance. Our dynamic parallel algorithm is validated with extensiveexperimental results.
|
[
"Thanks to the duality relationship between maxflow and minimal {{formula:2cd65ef6-69fc-4d52-94e2-d1c6d6f7d32f}} cut, the maxflow algorithms, usually categorized as augmenting-path-based {{cite:15fa182ff86e9466d4ad9241fa5e45dcfbc62477}}, {{cite:28ba28df0076d8d2d473d66a3774324401e12ede}}, {{cite:ee1618b61763413671113a497136cd4995e233cd}}, {{cite:5a235254fece1dcfaad79f8b7ff50949fa912f75}} and push-relabel-based {{cite:485114981dd52d2dc142d1f5f7a93c24fe256b24}}, {{cite:1713ea07eaa0acf61d437e152d3f8a7c1f81f023}}, {{cite:89a80de3c65a225eaaeefe858f2d58980a943428}}, {{cite:64500be394c0e379509470412ac5fcd929246c5b}}, have been a focus for parallelizing in the literature.",
"Push-relabel-based methods {{cite:485114981dd52d2dc142d1f5f7a93c24fe256b24}}, {{cite:1713ea07eaa0acf61d437e152d3f8a7c1f81f023}}, {{cite:89a80de3c65a225eaaeefe858f2d58980a943428}}, {{cite:64500be394c0e379509470412ac5fcd929246c5b}} are relatively easy to parallelize due to their memory locality feature. Therefore, a number of parallelized maxflow algorithms have been developed {{cite:e2b83f732602170c56a48f36fa50e8984192a2ed}}, {{cite:f548f1ee532de0cb0f71e10451ca0525403984be}}, {{cite:cf331d14d34860941fc588ba490addd7b9be3176}}, {{cite:f3475b83911cb96c2034f110e5ec36f8c2fcf3cd}}. These algorithms generally exhibit superior performance on huge 3D grids with high connectivity. Nevertheless, they are usually unable to achieve the same efficiency on 2D grids or moderately sized sparse 3D grids, which is common in many computer vision problems, as the state-of-the-art serial augmenting path method {{cite:28ba28df0076d8d2d473d66a3774324401e12ede}}, {{cite:5a235254fece1dcfaad79f8b7ff50949fa912f75}} could do on a commodity multi-core platform. The GPU implementation {{cite:7a943417907ff1a3d976418e4466acfea1df3734}} often fails to produce the correct results unless the amount of regularization is low {{cite:27c89d5aded6d2daf36ecd3f781b9d9004e9a1eb}}. The data transfer between the main memory and the graphics card is also a bottleneck for the GPU-based implementation.",
"Perhaps the most widely used graph cuts solver in the computer vision community is the algorithm proposed by Boykov and Kolmogorov {{cite:28ba28df0076d8d2d473d66a3774324401e12ede}} (called the BK-algorithm). This is a serial augmenting path maxflow algorithm that effectively reuses the two search trees originated from {{formula:f3475277-156a-425d-9d5a-6df9b854ff23}} and {{formula:fa7bd9b5-e530-4bb1-bb7a-f076af985934}} , respectively. To parallelize the BK-algorithm for further efficiency, the graph is usually split into multiple parts, either disjoint or overlapping. Liu and Sun {{cite:fc5b5c0d13757434cc4d2d0c4e7b89c0995de43b}} uniformly partitioned the graph into a number of disjoint subgraphs, concurrently ran the BK-algorithm in each subgraph to obtain short-range search trees within each subgraph, and then adaptively merged adjacent subgraphs to enlarge the search range until only one subgraph remained and all augmenting paths were found. For some 2D image segmentation cases, this algorithm can achieve a near-linear speedup with up to 4 computational threads. However, this method requires a shared-memory model, which makes it difficult to use on distributed platforms. To make the algorithm applicable to both shared and distributed models, Strandmark and Kahl {{cite:27c89d5aded6d2daf36ecd3f781b9d9004e9a1eb}}, {{cite:f981f90a93ff0ae4bade99cf0447d0ea9efa4da3}} proposed a new parallel and distributed BK-algorithm (called the parallel BK-algorithm), which splits the graph into overlapped subgraphs based on dual decomposition {{cite:bdbe37d9cd2820a4051d992d5b4ff6a1eaa97e9f}}, {{cite:a1a89769ca33093d5bfc7443a32f61733abff6e4}}. The BK-algorithm is then run in a parallel and iterative fashion. Unfortunately, due to the possible multiple optimal solutions of the BK-algorithm, which is an inherent characteristic of all graph cuts methods, the parallel BK-algorithm may fail to converge in some cases {{cite:54b1433a127aa064d6177ad7d11ee5e560bb795c}}.",
"By combining push-relabel and path augmentation methods, Shekhovtsov and Hlavac {{cite:54b1433a127aa064d6177ad7d11ee5e560bb795c}} proposed an algorithm to further reduce the number of message exchanges in the region push-relabel algorithm {{cite:f3475b83911cb96c2034f110e5ec36f8c2fcf3cd}}. Bhusnurmath and Taylor {{cite:4398f8c793897bbfd9f785d8b8c309dad24680d8}} reformulated graph cuts as an {{formula:770b5fc7-9f7b-4a9c-9c2e-c3ea454dabb2}} minimization problem and provided a highly parallelized implementation. However, even its GPU-based implementation is not significantly faster than the BK-algorithm for any type of graph."
] |
[] |
\"Uber die Klassifizierung von Knoten in dynamischen Netzwerken mit
Inhalt
|
This paper explains the DYCOS-Algorithm as it was introduced in by Aggarwaland Li in 2011. It operates on graphs whichs nodes are partially labeled andautomatically adds missing labels to nodes. To do so, the DYCOS algorithm makesuse of the structure of the graph as well as content which is assigned to thenode. Aggarwal and Li measured in an experimental analysis that DYCOS adds themissing labels to a Graph with 19396 nodes of which 14814 are labeled andanother Graph with 806635 nodes of which 18999 are labeld on one core of anIntel Xeon 2.5 GHz CPU with 32 G RAM within less than a minute. Additionally,extensions of the DYCOS algorithm are proposed. ----- In dieser Arbeit wird der DYCOS-Algorithmus, wie er 2011 von Aggarwal und Livorgestellt wurde, erkl\"art. Er arbeitet auf Graphen, deren Knoten teilweisemit Beschriftungen versehen sind und erg\"anzt automatisch Beschriftungen f\"urKnoten, die bisher noch keine Beschriftung haben. Dieser Vorgang wird"Klassifizierung" genannt. Dazu verwendet er die Struktur des Graphen sowietextuelle Informationen, die den Knoten zugeordnet sind. Die von Aggarwal undLi beschriebene experimentelle Analyse ergab, dass er auch auf dynamischenGraphen mit 19396 bzw. 806635 Knoten, von denen nur 14814 bzw. 18999beschriftet waren, innerhalb von weniger als einer Minute auf einem Kern einerIntel Xeon 2.5 GHz CPU mit 32 G RAM ausgef\"uhrt werden kann. Zus\"atzlich wirddie Ver\"offentlichung von Aggarwal und Li kritisch er\"ortert und und eswerden m\"ogliche Erweiterungen des DYCOS-Algorithmus vorgeschlagen.
|
[
"Sowohl das Problem der Knotenklassifikation, als auch das derTextklassifikation, wurde bereits in verschiedenen Kontexten analysiert. Jedochscheinen bisher entweder nur die Struktur des zugrundeliegenden Graphen odernur Eigenschaften der Texte verwendet worden zu sein.",
"So werden in {{cite:2d1733db890ac4ac75aaa5e103fb8c38607b8f3e}}, {{cite:ce85429dc907bf013ffd7bfa86846469c6465e81}} unter anderem Verfahren zurKnotenklassifikation beschrieben, die wie der in {{cite:1c1cf3731080d83152e644c5432b97681130da77}}vorgestellte DYCOS-Algorithmus, um den es in dieser Ausarbeitung geht, auch aufRandom Walks basieren.",
"Obwohl es auch zur Textklassifikation einige Paper gibt{{cite:dabaf1a9f7fae246ca7d9f001dc2c8c9af9b2235}}, {{cite:35f5388700a2179edb75a5aa585dc68b1a7bea2e}}, geht doch keines davon auf denSpezialfall der Textklassifikation mit einem zugrundeliegenden Graphen ein.",
"Die vorgestellten Methoden zur Textklassifikation variieren außerdem sehrstark. Es gibt Verfahren, die auf dem bag-of-words-Modell basieren{{cite:18c22e92a46f86e414f7e9c3ff0a6c418727fe72}} wie es auch im DYCOS-Algorithmus verwendetwird. Aber es gibt auch Verfahren, die auf demExpectation-Maximization-Algorithmus basieren {{cite:3896670ac484ccb3a9bcfad9909d134ae71b06ee}}oder Support VectorMachines nutzen {{cite:b7929d16f2e4770d77dc400f82a6f549b2ee0cfb}}.",
"Es wäre also gut Vorstellbar, die Art und Weise wie die Texte in dieKlassifikation des DYCOS-Algorithmus einfließen zu variieren. Allerdings istdabei darauf hinzuweisen, dass die im Folgenden vorgestellte Verwendung derTexte sowohl einfach zu implementieren ist und nur lineare Vorverarbeitungszeitin Anzahl der Wörter des Textes hat, als auch es erlaubt einzelne Knoten zuklassifizieren, wobei der Graph nur lokal um den zu klassifizierenden Knotenbetrachten werden muss."
] |
[] |
A Lambda-Calculus Foundation for Universal Probabilistic Programming
|
We develop the operational semantics of an untyped probabilisticlambda-calculus with continuous distributions, as a foundation for universalprobabilistic programming languages such as Church, Anglican, and Venture. Ourfirst contribution is to adapt the classic operational semantics oflambda-calculus to a continuous setting via creating a measure space on termsand defining step-indexed approximations. We prove equivalence of big-step andsmall-step formulations of this distribution-based semantics. To move closer toinference techniques, we also define the sampling-based semantics of a term asa function from a trace of random samples to a value. We show that thedistribution induced by integrating over all traces equals thedistribution-based semantics. Our second contribution is to formalize theimplementation technique of trace Markov chain Monte Carlo (MCMC) for ourcalculus and to show its correctness. A key step is defining sufficientconditions for the distribution induced by trace MCMC to converge to thedistribution-based semantics. To the best of our knowledge, this is the firstrigorous correctness proof for trace MCMC for a higher-order functionallanguage.
|
[
"To the best of our knowledge, the only previous theoreticaljustification for trace MCMC is the recent work by {{cite:2575087ecaff107b5076d2b1b485dfdbb20006da}},who show correctness of trace MCMC for the imperative probabilistic language R2 {{cite:212f33653f4a3214f7bf39bd1e00f410cd415d8c}}.Their result does not apply to higher-order languages such as Church or our {{formula:79e702f0-1406-4b70-8374-15053832f14b}} -calculus, nor to programs that do not almost surely terminate {{cite:3afe2e87472cb815b505011d894030450a5feab7}}.Their algorithm is different from ours in that it exploits the explicit storage locations found in imperative programs, keeping one sample trace per location.The authors do state that the space of traces in their language isequipped with a “stock” measure, and that the distributions ofprogram traces and transitions are given by densities with respect tothat measure. They do not, however, show that these densities aremeasurable.Their proof of correctness only shows that the acceptance ratio{{formula:7e133b9f-cf23-4947-a868-b5863bc0aaf9}} computed by their algorithm matches Hasting's formula: theauthors do not prove convergence of the resulting Markov chain.Indeed, properties such as irreducibility and aperiodicity depend onthe choices of parameters in the algorithm.",
"Other probabilistic language implementations also use trace MCMC inference,including Church {{cite:f25d7a384420295740e5200d5a6e2cd72884dba6}},Venture {{cite:c53b14b6686b234bc3e9f99d87b1d6c385e749dc}}, WebPPL {{cite:5c79d8760103346689b0cde5f41f0244166e31da}},and Anglican {{cite:c7fb70c8100813879551ba4e8d9a8508c78d7c7a}}.These works focusing on efficiency and convergence properties,and do not state formal correctness claims for their implementations.",
"{{cite:50714bca89269d13ef170016673586784240b054}} give a general program transformationfor a probabilistic language to support trace MCMC, with a focus onlabelling sample points in order to maximise sample reuse. Extendingour trace semantics with such labelling is important future work,given that {{cite:a353c2d8aa4de1a6344919d02646922573ca2b97}} points out some difficultieswith the transformation and proposes alternatives.",
"Many recent probabilistic languages admit arbitrary non-negativescores. This is done either by having an explicit{{formula:e77ffcbd-ec44-49d2-996a-3b6ab2002004}} -like function, as inWebPPL (called factor), or byobserving that a particular value {{formula:c41178fc-b98e-4471-932f-6ce4638f9f92}} was drawn from a givendistribution {{formula:63626c90-25ed-4837-ab93-a05ca8153c7f}} (without adding it to the trace), asin Web Church (written {{formula:c8ea63ae-ff7f-4dc2-8a35-c5256f33a159}} ) or Anglican (observe {{formula:48f5e727-8981-4981-813f-9bb0612f61ff}} ).In recent work for a non-recursive {{formula:2f94fb3d-e6e5-44ee-bdae-f5c5776805d1}} -calculus withscore, {{cite:5720115b50ba7b97a7e9bc0e05e72ef119b5c758}} note thatunbounded scores introduce the possibility of “infinite modelevidence errors”. As seen in Section REF ,even statically bounded scores exhibit this problem in the presence of recursion.",
"{{cite:3796d4d21bdc27550a994473ba759a323337c373}} gives a semantics of imperativeprobabilistic programs as partial measurable functions from infiniterandom traces to final states, which serves as the model for our tracesemantics. Kozen also proves this semantics equivalent to adomain-theoretic one.{{cite:50a778e6ffa3bec484123fc998d1c05ed36429b7}} give an operational version of Kozen'strace-based semantics for a {{formula:631a11d5-e5f7-4c4c-993e-e2db41e70ecd}} -calculus with recursion, but“do not investigate measure-theoretic properties”.{{cite:3a96196ec34dd325ac410060c893aeb0ebb3be10}} generalise Kozen's trace-basedsemantics to consider probabilistic programs as measurable functionsfrom a probability space into a semantics domain, and study abstractinterpretation in this setting.{{cite:a1a17960b58037fd82f5c15c2c7ee04904e07f36}} use a pre-image version of Kozen'ssemantics to obtain an efficient implementation using rejectionsampling.{{cite:5fca5ad9938a5f29dc1450c7ec387833f18dfcd8}} define a monadic embedding ofprobabilistic programming in Haskell along the lines of Kozen'ssemantics; their paper describes various inference algorithms but hasno formal correctness results.",
"{{cite:e694476b2d2f69686ded4ce959024ddaab4545ed}} provide a monadic denotationalsemantics for a first-order functional language with discreteprobabilistic choice, and a Haskell implementation of the expectationmonad using variable elimination.{{cite:284288ff29a04f5ced7795391ed8b65412553162}} define a denotational semantics based ondensity functions for a restricted first-order language withcontinuous distributions. They also present a type system ensuringthat a given program has a density.",
"{{cite:433db4de8bcf79d4f81e0dbfbcacc7f54ea9f575}} defines operational anddomain-theoretic semantics for a {{formula:3e19e240-ca59-4894-ab80-487bb12ad346}} -calculus with discreteprobabilistic choice and a fixpoint construct. Our distribution-basedoperational semantics generalises Jones's to deal with continuousdistributions.Like Kozen's and Jones's semantics, our operational semantics makesuse of the partially additive structure on the category ofsub-probability kernels {{cite:fe4179b602dad9f91d29237439247b209421dca1}} in order totreat programs that make an unbounded number of random draws.{{cite:5720115b50ba7b97a7e9bc0e05e72ef119b5c758}} give a domain-theoreticsemantics for a {{formula:7be3770e-7626-4bdc-845b-c2c636c53852}} -calculus with continuous distributions andunbounded score, but without recursion.While giving a fully abstract domain theory for probabilistic{{formula:5e081c16-c736-439d-80d8-057c97b341ce}} -calculi with recursion is known to behard {{cite:88ecc2dbe18b7d7debf70e75783dc3e277293fab}}, there have been recent advances using probabilistic coherencespaces {{cite:780b8e1b6cc64f04f2459c67a6d67c36233cce7d}}, {{cite:f61fa2d327926fad6d4fdd41c691eee68ac057d6}} and gamesemantics {{cite:357bbc67bde33e1af941f1006186312d24b5beb5}}, which in some cases are fullyabstract. We see no strong obstacles in applying any of these to atyped version of our calculus, but it is beyond the scope of thiswork.Another topic for future work are methodologies for equivalencechecking in the style of logical relations or bisimilarity, which havebeen recently shown to work well in discrete probabilisticcalculi {{cite:df93b653af2860baf11e03d1fd4b95110ef573a8}}."
] |
[] |
Deep Learning over Multi-field Categorical Data: A Case Study on User
Response Prediction
|
Predicting user responses, such as click-through rate and conversion rate,are critical in many web applications including web search, personalisedrecommendation, and online advertising. Different from continuous raw featuresthat we usually found in the image and audio domains, the input features in webspace are always of multi-field and are mostly discrete and categorical whiletheir dependencies are little known. Major user response prediction models haveto either limit themselves to linear models or require manually building uphigh-order combination features. The former loses the ability of exploringfeature interactions, while the latter results in a heavy computation in thelarge feature space. To tackle the issue, we propose two novel models usingdeep neural networks (DNNs) to automatically learn effective patterns fromcategorical feature interactions and make predictions of users' ad clicks. Toget our DNNs efficiently work, we propose to leverage three featuretransformation methods, i.e., factorisation machines (FMs), restrictedBoltzmann machines (RBMs) and denoising auto-encoders (DAEs). This paperpresents the structure of our models and their efficient training algorithms.The large-scale experiments with real-world data demonstrate that our methodswork better than major state-of-the-art models.
|
[
"Click-through rate, defined as the probability of the ad click from a specific user on a displayed ad, is essential in online advertising {{cite:677bf2a4592aa6cff673590968819c682813d397}}. In order to maximise revenue and user satisfaction, online advertising platforms must predict the expected user behaviour for each displayed ad and maximise the expectation that users will click. The majority of current models use logistic regression based on a set of sparse binary features converted from the original categorical features via one-hot encoding {{cite:253eb49b8a9c67a2d88f1fd90bc8245397ed0542}}, {{cite:b7430d51c8e3a49a058639cecde0dcd88283503f}}. Heavy engineering efforts are needed to design features such as locations, top unigrams, combination features, etc. {{cite:84670513374d24c05d040f7bb816689445980617}}.",
"Embedding very large feature vector into low-dimensional vector spaces is useful for prediction task as it reduces the data and model complexity and improves both the effectiveness and the efficiency of the training and prediction. Various methods of embedding architectures have been proposed {{cite:3d8d004cced683e7dda1a1e0ecd17b748f0fdff8}}, {{cite:976a69a72b9eaa73b19710ec4949205ef10e6f81}}.Factorisation machine (FM) {{cite:6c2ec6f568573e4c6af258e076afc36a430fa6aa}}, originally proposed for collaborative filtering recommendation, is regarded as one of the most successful embedding models. FM naturally has the capability of estimating interactions between any two features via mapping them into vectors in a low-rank latent space.",
"Deep Learning {{cite:528a61bbe13e649ffaed5a34d63d5e9daaea9f6d}} is a branch of artificial intelligence research that attempts to develop the techniques that will allow computers to handle complex tasks such as recognition and prediction at high performance. Deep neural networks (DNNs) are able to extract the hidden structures and intrinsic patterns at different levels of abstractions from training data. DNNs have been successfully applied in computer vision {{cite:5989c726734001de1914fc58932cead7cd4a2c66}}, speech recognition {{cite:0f3b29b24f0aeb77d492f5010b48d2046cff83c2}} and natural language processing (NLP) {{cite:bb34a9932a371d6cc42fc71086eebefb2920a878}}, {{cite:162fe99beac90652fa90389cba694fdaab519886}}, {{cite:82736d03be0fb504b7f2bea22e011d42d9246a1f}}. Furthermore, with the help of unsupervised pre-training, we can get good feature representation which guides the learning towards basins of attraction of minima that support better generalisation from the training data {{cite:67d17bc87bcabbc41efa60ef1fad7e7f89f1724f}}. Usually, these deep models have two stages in learning {{cite:35c64e5d116b31144ca17bc8c58fe8606aa04c58}}: the first stage performs model initialisation via unsupervised learning (i.e., the restricted Boltzmann machine or stacked denoising auto-encoders) to make the model catch the input data distribution; the second stage involves a fine tuning of the initialised model via supervised learning with back-propagation. The novelty of our deep learning models lies in the first layer initialisation, where the input raw features are high dimensional and sparse binary features converted from the raw categorical features, which makes it hard to train traditional DNNs in large scale. Compared with the word-embedding techniques used in NLP {{cite:162fe99beac90652fa90389cba694fdaab519886}}, {{cite:82736d03be0fb504b7f2bea22e011d42d9246a1f}}, our models deal with more general multi-field categorical features without any assumed data structures such as word alignment and letter-n-gram etc."
] |
[] |
nodewatcher: A Substrate for Growing Your own Community Network
|
Community networks differ from regular networks by their organic growthpatterns -- there is no central planning body that would decide how the networkis built. Instead, the network grows in a bottom-up fashion as more peopleexpress interest in participating in the community and connect with theirneighbours. People who participate in community networks are usually volunteerswith limited free time. Due to these factors, making the management ofcommunity networks simpler and easier for all participants is the key componentin boosting their growth. Specifics of individual networks often forcecommunities to develop their own sets of tools and best practices which arehard to share and do not interoperate well with others. We propose a newgeneral community network management platform nodewatcher that is built aroundthe core principle of modularity and extensibility, making it suitable forreuse by different community networks. Devices are configured usingplatform-independent configuration which nodewatcher can transform intodeployable firmware images, eliminating any manual device configuration,reducing errors, and enabling participation of novice maintainers. An embeddedmonitoring system enables live overview and validation of the whole communitynetwork. We show how the system successfully operates in an actual communitywireless network, wlan Slovenija.
|
[
"A lot of research has been done into individual wireless mesh network building blocks. Among them are routing {{cite:0ef7dffae7563180b9be235e81909bc7dc64c6ee}}, {{cite:b3df04fffaec70bdafd9f0d3e7704e168c08863c}}, {{cite:e1cfd0293d872c5969aceb81dfc2a34f1ac273c8}}, security {{cite:4fe3f86d7568a2671a9aa0ae65082243164da5d2}}, and analyses of topologies, performance, mobility {{cite:75f29109979496eced48e830e23e0113ff97fe6a}}, {{cite:eca73a388f89bcef1ed6efa4b6b337dfd40de451}}, {{cite:7d3c8f92bbdb9876342c3140590b720c79c0dd0d}}, {{cite:f30916c744117bb0e5dd19321bd70ee6102a299d}}, {{cite:f963cdab06cbd63446a1119a69416761d7ee38d4}}.But research into community network management solutions and best practices still remains scarce.This is why most of the related work in this area comes from the individual community networks which have each developed its own solutions, practices and philosophy.",
"In this section we survey the most visible network management solutions, traditional ones and specialized solutions developed by community networks worldwide.We compare them to the approach taken by nodewatcher platform."
] |
[] |
Implicit Look-alike Modelling in Display Ads: Transfer Collaborative
Filtering to CTR Estimation
|
User behaviour targeting is essential in online advertising. Compared withsponsored search keyword targeting and contextual advertising page contenttargeting, user behaviour targeting builds users' interest profiles viatracking their online behaviour and then delivers the relevant ads according toeach user's interest, which leads to higher targeting accuracy and thus moreimproved advertising performance. The current user profiling methods includebuilding keywords and topic tags or mapping users onto a hierarchical taxonomy.However, to our knowledge, there is no previous work that explicitlyinvestigates the user online visits similarity and incorporates such similarityinto their ad response prediction. In this work, we propose a general frameworkwhich learns the user profiles based on their online browsing behaviour, andtransfers the learned knowledge onto prediction of their ad response.Technically, we propose a transfer learning model based on the probabilisticlatent factor graphic models, where the users' ad response profiles aregenerated from their online browsing profiles. The large-scale experimentsbased on real-world data demonstrate significant improvement of our solutionover some strong baselines.
|
[
"Ad Response Prediction aims at predicting the probability that a specific user will respond (e.g., click) to an ad in a given context {{cite:69af3a367f3ceb5c8267e24bd182c8ad71a79dfd}}, {{cite:6f19823c06ef1539656d94e555bb4b81d7e27e3d}}. Such context can be either a search keyword {{cite:01f8494a778ccac1f5e5cb0310a47b65dac32e10}}, webpage content {{cite:2f4e0487ee46267cb0d16025a728aae7e3da0e36}}, or other kinds of real-time information related to the underlying user {{cite:ecaabe34d99de4105880ed4390b7b879aad7f74f}}. From the modelling perspective, many user response prediction solutions are based on linear models, such as logistic regression {{cite:aa77132002758966d3337ccce4a3b10337009bbb}}, {{cite:939c94f987a71d1c8c163647925428f4061726bc}} and Bayesian probit regression {{cite:01f8494a778ccac1f5e5cb0310a47b65dac32e10}}. Despite the advantage of high learning efficiency, these linear models suffer from the lack of feature interaction and combination {{cite:8e4dda26184984bff3ce3409e95e26a8acd30e29}}. Thus non-linear models such as tree models {{cite:8e4dda26184984bff3ce3409e95e26a8acd30e29}} and latent vector models {{cite:f7fb2adadb6c74f3f57cc1bd7727947a451b09bd}}, {{cite:bf659a4ddafff8b73e873387cb79ac1180dc691a}} are proposed to catch the data non-linearity and interactions between features. Recently the authors in {{cite:e1c0812c10d7fa3c547fefc144aac3fafd725a4b}} proposed to first learn combination features from gradient boosting decision trees (GBDT) and, based on the tree leaves as features, learn a factorisation machine (FM) {{cite:5f0417b3bffa090d32b235716b3e3efe603a6f18}} to build feature interactions to improve ad click prediction performance.",
"Collaborative Filtering (CF) on the other hand is a technique for personalised recommendation {{cite:2200a2628b39ee6254a55684f4e625e28fb3ab13}}. Instead of exploring content features, it learns the user or/and item similarity based on their interactions. Besides the user(item)-based approaches {{cite:d4d050fdd582c844dd4e0c1da4b6799dd31a1881}}, {{cite:ad27b0065074788bfcf69245ccadffc7867fb606}}, latent factor models, such as probabilistic latent semantic analysis {{cite:8b325c745a4f43b22a4ad5f394c4e1ece90b3e3b}}, matrix factorisation {{cite:539a0b07b25c34d1a02f7b877e92bd6fc2e6b089}} and factorisation machines {{cite:5f0417b3bffa090d32b235716b3e3efe603a6f18}}, are widely used model-based approaches. The key idea of the latent factor models is to learn a low-dimensional vector representation of each user and item to catch the observed user-item interaction patterns. Such latent factors have good generalisation and can be leveraged to predict the users' preference on unobserved items {{cite:539a0b07b25c34d1a02f7b877e92bd6fc2e6b089}}. In this paper, we explore latent models of collaborative filtering to model user browsing patterns and use them to infer users' ad click behaviour.",
"Transfer Learning deals with the learning problem where the learning data of the target task is expensive to get, or easily outdated, via transferring the knowledge learned from other tasks {{cite:02aa8230a529707e0089a80e48aa3265272d1280}}. It has been proven to work on a variety of problems such as classification {{cite:82f4912fd09ca6efb0c6a2f064ca2071678c5d2e}}, regression {{cite:4b171918027f64d95737fab7876e04ef3cbcc2eb}} and collaborative filtering {{cite:0043aad2daefcb8df059201873717d5b5c660fa4}}. Different from multi-task learning, where the data from different tasks are assumed to drawn from the same distribution {{cite:536113acf366d54517756b42194cb3dc44d1885b}}, transfer learning methods may allow for arbitrary source and target tasks. In online advertising field, the authors in a recent work {{cite:0a2d0326b1706c3d15da427c7b636976eba85abb}} proposed a transfer learning scheme based on logistic regression prediction models, where the parameters of ad click prediction model were restricted with a regularisation term from the ones of user web browsing prediction model. In this paper, we consider it as one of the baselines."
] |
[] |
Wigner Functions for the Pair Angle and Orbital Angular Momentum
|
The problem of constructing physically and mathematically well-defined Wignerfunctions for the canonical pair angle and angular momentum is solved. While akey element for the construction of Wigner functions for the planar phase space(q,p) in R^2 is the Heisenberg-Weyl group, the corresponding group for thecylindrical phase space (theta,p) in S^1 x R is the Euclidean group E(2) of theplane and its unitary representations. Here the angle theta is replaced by thepair (cos theta, sin theta) which corresponds uniquely to the points on theunit circle. The main structural properties of the Wigner functions for theplanar and the cylindrical phase spaces are strikingly similar. A crucial roleplays the sinc function which provides the interpolation for the discontinuousquantized angular momenta in terms of the continuous classical ones, inaccordance with the famous Whittaker cardinal function, well-known frominterpolation and sampling theories. The quantum mechanical marginaldistributions for angle (continuous) and angular momentum (discontinuous) areas usual uniquely obtained by appropriate integrations of the (theta,p) Wignerfunction. Among the examples discussed is an elementary system of simple "cat"states.
|
[
"The approach discussed above for phase spaces of the topological type {{formula:e1e376a2-81da-4323-99e0-e7b98b8244eb}} canbe generalized, e.g. to a free rigid body in 3 dimensions with one point fixed. Its configuration spacecan be identified with the group {{formula:43b46a50-36ac-4cd6-bd72-be2ee29dc9ff}} {{cite:135226ca75142274139446dd1112efb104f6e7a3}}, {{cite:e36cebe3c0be22095b1e37ce0cf7c98cffdac9c8}}. Its 2-fold covering {{formula:ba8f552f-acac-4e77-8831-44fc4e435257}} has the topology of {{formula:79a8109a-3939-47ba-bd1c-e85cfdc5e53b}} .The associated phase space {{formula:07a7dab5-793f-4eb4-8a14-afeae597060f}} can be quantized in terms of unitary representationsof the Euclidean group {{formula:4dadcc8a-1eed-4fce-9cdb-4d2b94922409}} {{cite:25efe4095b4e65b7396258ecf0f73f8bbecc3af1}}.",
"At a late stage of the present investigations I became aware of the work by Leaf {{cite:5a921967035d2f7bf9ac4f3febd02c43f6a237fa}}, {{cite:a021409ad0d88fcae471b25fc62f6f72c5a92fbb}} in which the operator{{formula:265d5cb6-24e4-4a94-bf65-da3d0be8934a}} plays a role with respect to the {{formula:37c74ea2-4f78-47c1-8710-0a97236b5626}} Wigner function which corresponds to that of the operator/matrix{{formula:248f7a73-fd32-432c-9199-fd6cd6100b88}} in the discussions above. Leaf's approach was used by de Groot and Suttorp in their textbook {{cite:977b630288dbe93cada1dd9b4cd8101e3692ce39}}.",
"A Wigner-Moyal function on the circle very similar in structure to the one in Eq. (REF ) above can be foundin Refs. {{cite:c79895036e35c20075bc861b1adec4669f219cd8}}, {{cite:1e7bd5f93a2702875ca705f9fb61f7e6fc47830b}}. There, however, the classical angular momentum is treated as discontinuous, contrary to its propertiesin its classical phase space and its treatment in the present paper!",
"I am again very grateful to the DESY TheoryGroup for its continuous generous hospitality since my retirementfrom the Institute for Theoretical Physics of the RWTH Aachen.",
"I thank Krzysztof Kowalski for drawing my attention first to Ref. {{cite:c79895036e35c20075bc861b1adec4669f219cd8}} after a shorter Letter versionof the present paper had appeared (arXiv:1601.02520v2) and to Ref. {{cite:1e7bd5f93a2702875ca705f9fb61f7e6fc47830b}} after the completion of the present paper.I thank Hartmann Römer for a critical reading of the manuscriptand the suggestion to include thermal states among the examples in Sec. IV.Finally I am indebted to my son David for providing the figures and for his help with the final typewriting."
] |
[] |
Git4Voc: Git-based Versioning for Collaborative Vocabulary Development
|
Collaborative vocabulary development in the context of data integration isthe process of finding consensus between the experts of the different systemsand domains. The complexity of this process is increased with the number ofinvolved people, the variety of the systems to be integrated and the dynamicsof their domain. In this paper we advocate that the realization of a powerfulversion control system is the heart of the problem. Driven by this idea and thesuccess of Git in the context of software development, we investigate theapplicability of Git for collaborative vocabulary development. Even thoughvocabulary development and software development have much more similaritiesthan differences there are still important differences. These need to beconsidered within the development of a successful versioning and collaborationsystem for vocabulary development. Therefore, this paper starts by presentingthe challenges we were faced with during the creation of vocabulariescollaboratively and discusses its distinction to software development. Based onthese insights we propose Git4Voc which comprises guidelines how Git can beadopted to vocabulary development. Finally, we demonstrate how Git hooks can beimplemented to go beyond the plain functionality of Git by realizingvocabulary-specific features like syntactic validation and semantic diffs.
|
[
"Collaborative vocabulary development is an active research area in the Semantic Web community {{cite:709710dfde1fe3f84121ec8cc8aa93b74d26be24}}.Existing approaches like WebProtégé {{cite:043a2c56d94cf5722acdb766cf3a8edad69737f3}} provides a collaborative web frontend for a subset of the functionality of the Protégé OWL editor.The aim of WebProtégé, is to lower the threshold for collaborative ontology development.Neologism {{cite:6ccedcb3523e9ae482d226fdbc77e34e975687fc}} is a vocabulary publishing platform, with a focus on ease of use and compatibility with Linked Data principles.Neologism focuses more on vocabulary publishing and less on collaboration.VocBench {{cite:393c0affa99d1f1d9767b80eb79b45cb80e34de5}}, is an open source web application for editing thesauri complying with the SKOS and SKOS-XL standards.VocBench has a focus on collaboration, supported by workflow management for content validation and publication.",
"The main limitation of the aforementioned tools is the lack of version control.Therefore, we only consider approaches focused on using version control systems for collaborative vocabulary development.",
"SVoNt {{cite:a5302b93692bb4853c4f8cde6b445696b482f5ab}} extends the functionality of Apache SubVersion (SVN) by providing a possibility for versioning OWL conform lightweight description logic.SVN manipulates only with deltas of files, therefore SVoNt use a separate server to create conceptual changes between versions of ontologies.These changes are generated as a result of diff operation between the modified ontology and the base ontology.ContentCVS {{cite:078fe4d1e10138342eb5df130ca538b27506f595}} is a Protégé plugin.It adapts concepts from concurrent versioning to enable developers to work in parallel.Moreover, it has features for conflict detection and resolution by checking structure and semantic of the ontology versions.In {{cite:8a41b31a277dd4afe4875da2dce5ea139eb873fb}} is described how the developers of RDA Vocabularieshttp://www.rdaregistry.info adopt rules from SemVerhttp://www.semver.org to realize a meaningful versioning using Git.Additionally, it provides general notes for organizing the vocabulary development in branches.{{cite:3309103b20182285f2d671c59ee2530ee7ecce9e}} describes Owl2VCS, a toolset designed to facilitate version control of OWL 2 ontologies using version control systems.It can be integrated as an external tool with Git, Mercurial and Subversion and provide algorithms for structural diff {{cite:58ddbac430503ecf2ac14cf920dd32d3944786eb}}.However, none of the above mentioned approaches cover all the identified requirements (c.f. Requirements) for collaborative vocabulary development.On the contrary, our work analyze and address each one of them by using Git and Git4Voc as an extension."
] |
[] |
Learning Subclass Representations for Visually-varied Image
Classification
|
In this paper, we present a subclass-representation approach that predictsthe probability of a social image belonging to one particular class. We explorethe co-occurrence of user-contributed tags to find subclasses with a strongconnection to the top level class. We then project each image on to theresulting subclass space to generate a subclass representation for the image.The novelty of the approach is that subclass representations make use of notonly the content of the photos themselves, but also information on theco-occurrence of their tags, which determines membership in both subclasses andtop-level classes. The novelty is also that the images are classified intosmaller classes, which have a chance of being more visually stable and easierto model. These subclasses are used as a latent space and images arerepresented in this space by their probability of relatedness to all of thesubclasses. In contrast to approaches directly modeling each top-level classbased on the image content, the proposed method can exploit more informationfor visually diverse classes. The approach is evaluated on a set of $2$ millionphotos with 10 classes, released by the Multimedia 2013 Yahoo! Large-scaleFlickr-tag Image Classification Grand Challenge. Experiments show that theproposed system delivers sound performance for visually diverse classescompared with methods that directly model top classes.
|
[
"Our work is closely related to method used for predicting tags with high intra-class variation, in particular, sub-category based methods or methods based on learning high level representations are often explored. We discuss these methods here in turn.",
"Generating sub-categories has been considered as an effective method to deal with classification problems where intra-class variation is high. The ImageNet {{cite:9967e17f2c6e059f0b098ea3e73703201ecb0afe}} organizes image dataset with labels corresponding to a semantic hierarchy. This method is able to build comprehensive ontology for large scale dataset. However, for a particular dataset, the sub-categories generated by data driven strategies are expected to be more discriminative. {{cite:9c8c130644db0fad95dd2715c7bbff5baf0a47e5}} exploits co-watch information to learn latent sub-tags for video tag prediction. {{cite:557ff3fc1e8bba454db3365877b0a68cb1b75b82}} proposes to discover the image hierarchy by using both visual and tag information.Our method generates category-specific subclasses by exploring image/tag co-occurrence, and trains classifiers for each subclass-tag. These subclasses based models are expected to be discriminative in terms of estimating the target tags, corresponding to top-level classes.",
"Learning higher level representation is adopted when the low-level image features are not discriminative enough for the purpose of classification {{cite:cfc9b8d74aa8867203f4dd475508a8f243c8e16b}}, {{cite:e67ff6a201933e759b0d058d89fc2abef97d8347}}. For the supervised representation learning methods, a predefined set of models are trained based on image features. The output of these models is considered as the high level representations for predicting image categories. Recently, deep neural networks have been used on unsupervised learning image representations with large scale image dataset {{cite:79df64fdcbf4a8f4b18a4d1b97498bd36666bcb2}}, {{cite:98b9102c847abc4a437fd0fca20ed375358f6031}}. This representation has achieved promising results on different classification and tagging tasks {{cite:db0fe6121799a74985f2cfb5490e9a00f2780c01}}. Our methods used the output of trained subclass classifiers as the high level features. This structure can be easily extend to deeper levels by finding discriminative tags for the subclasses."
] |
[] |
Quantification of Ultrasonic Texture heterogeneity via Volumetric
Stochastic Modeling for Tissue Characterization
|
Intensity variations in image texture can provide powerful quantitativeinformation about physical properties of biological tissue. However, tissuepatterns can vary according to the utilized imaging system and areintrinsically correlated to the scale of analysis. In the case of ultrasound,the Nakagami distribution is a general model of the ultrasonic backscatteringenvelope under various scattering conditions and densities where it can beemployed for characterizing image texture, but the subtle intra-heterogeneitieswithin a given mass are difficult to capture via this model as it works at asingle spatial scale. This paper proposes a locally adaptive 3Dmulti-resolution Nakagami-based fractal feature descriptor that extendsNakagami-based texture analysis to accommodate subtle speckle spatial frequencytissue intensity variability in volumetric scans. Local textural fractaldescriptors - which are invariant to affine intensity changes - are extractedfrom volumetric patches at different spatial resolutions from voxellattice-based generated shape and scale Nakagami parameters. Using ultrasoundradio-frequency datasets we found that after applying an adaptive fractaldecomposition label transfer approach on top of the generated Nakagami voxels,tissue characterization results were superior to the state of art. Experimentalresults on real 3D ultrasonic pre-clinical and clinical datasets suggest thatdescribing tumor intra-heterogeneity via this descriptor may facilitateimproved prediction of therapy response and disease characterization.
|
[
"One of the effective statistical techniques used for modeling various backscattering conditions in medical ultrasound is the Nakagami distribution. This probabilistic distribution is known for its analytical simplicity and effective modeling of dense scatterers, accounting for amplitude and spacing, and can be reduced to a Rayleigh distribution under certain assumptions of scatterer density and number {{cite:65feef9a97071ef16df35dce3c1c49de234cdd16}}. Shankar et al. first proposed the Nakagami distribution for characterizing conditions ranging from pre- to post-Rayleigh existing in ultrasound images, and later for modeling the radio-frequency (RF) envelope of the ultrasound backscattered signal in characterizing B-mode breast masses {{cite:03b0ac95466e7b89cd5c817786f050e91ff3e308}}. Others have attempted to tackle the issue of accurate estimation of the Nakagami distribution. For instance {{cite:63ee72732c4a0896ded0ade3ae0e6eb27856dd41}} employed Gamma kernel density estimation to compute a smooth estimate of a distribution within small windows of B-mode ultrasound images, but the mixture of distributions occurring at the boundaries between structures was not accounted for. The impact of morphological parameters and tumor structures on the Nakagami parameters statistics were analyzed in {{cite:42714a97de5ea5b29817a5470cd8045d4eea3826}}. A limitation observed was that there was a need for a robust algorithm to compute the Nakagami parametric images that better delineate the structures and the context in and around the tumor. Characterizing homogeneous tissues via improving the smoothness of the Nakagami parametric images was shown in {{cite:fcdf732ff1ed683ffb6405c4942253ab8fc6a455}}. The technique relies on summing and averaging the Nakagami images formed using sliding windows with varying window sizes related to the transducer pulse length. However, a relatively large window size (required for stabilization) may affect the reliability of the estimated Nakagami parameters, and hence degrade the spatial resolution of the resulting Nakagami image.",
"The Nakagami distribution has been further employed as an image feature in various image analysis contexts. For instance, five contour features and the two Nakagami parameters were used for classification of benign and malignant breast tumors in {{cite:f225c9848e2c45816168e566001b9934f075e3dd}}. In a subsequent work malignant tumors were shown to be more pre-Rayleigh distributed than those from benign counterparts {{cite:c5ac52c9f7dddf25011bf042f00d35012e49d410}}; however, the calculation of the average intensity value in the Nakagami image makes it susceptible to spatial frequency intensity variability. Further, that particular technique was optimized for 2D ultrasound images which may not reliably represent heterogeneous distributions of scatterers (or speckle) encountered within a tumor volume. A Random Forest based solution to learn tissue-specific ultrasonic backscattering and a signal confidence for predicting heterogeneous composition in atherosclerotic plaques was proposed in {{cite:462a45c8209115adb2be8a1ac3971752263307d0}}. That technique was developed for intravascular ultrasound and risk assessment of plaque rupture {{cite:87563e0c3a87fa7977dbf1c0da33c30b34b8e455}}. Necrotic core was not considered in that method. Finally, {{cite:6fd14d11a0847aa59ed57c9f211a1cf3defa0856}}, {{cite:b53252ff94b8a148fb143791111dbc5c8f5b9d7b}} describe a Markov random field model combined with Nakagami distribution estimation to differentiate malignant melanoma from normal tissue. However it was found that the estimated scale model parameter was highly sensitive to image quality, and hence subtle variations could go unnoticed. For an overview of ultrasound tissue characterization we refer the reader to {{cite:c43ff74ac22f33599f7a00bd69e04a9e1014db4e}}. Previous work on ultrasound texture analysis of tumors has considered both global and local non-uniformity quantification of the tumor texture at only a single analysis scale. Herein we are primarily concerned with tumor intra-heterogeneity (i.e. micro-structures within the tumor speckle texture) which is more challenging."
] |
[] |
JADE, TraSMAPI and SUMO: A tool-chain for simulating traffic light
control
|
Increased stress, fuel consumption, air pollution, accidents and delays aresome of the consequences of traffic congestion usually incurring in tremendouseconomic impacts, which society aims to remedy in order to leverage asustainable development. Recently, unconventional means for modeling andcontrolling such complex traffic systems relying on multi-agent systems havearisen. This paper contributes to the understanding of such complex and highlydynamic systems by proposing an open-source tool-chain to implementmulti-agent-based solutions in traffic and transportation. The proposedapproach relies on two very popular tools in both domains, with focus ontraffic light control. This tool-chain consists in combining JADE (Java AgentDEvelopment Framework), for the implementation of multi-agent systems, withSUMO (Simulation of Urban MObility), for the microscopic simulation of trafficinteractions. TraSMAPI (Traffic Simulation Manager Application ProgrammingInterface) is used to combine JADE and SUMO allowing communication betweenthem. A demonstration of the concept is presented to illustrate the mainfeatures of this tool-chain, using Q-Learning as the reinforcement learningmethod for each traffic light agent in a simulated network. Results demonstratethe feasibility of the proposed framework as a practical means to experimentwith different agent-based designs of intelligent transportation solutions.
|
[
"The specific case of traffic lights is one of the areas where much has been researched for new solutions, from the design of intersections {{cite:daf57312c96eaf7ea25dabea27879f9010b3944a}} (including physical layout and semaphore plans), to the definition of semaphore plans through statistical analysis. Current solutions try to answer the highly dynamic system {{cite:2e6f318fb2a2639018fdd83d5e5c5bebfe349d5c}} using coordinated control. Several methodologies have been used such as genetic algorithms {{cite:3615f5a16262c1d1245fa4703d92f0bd0218159f}}, fuzzy logic {{cite:71e79c627804680c860d11a2710a46e9e77aff9b}} and reinforcement learning {{cite:6a7c2c35824d8321862b9d1bd341c122096ee20c}}.",
"To date, there are not many solutions for traffic that make full use of the intelligent agent concept. However, the multi-agent system approach has become recognized as a convenient approach for modelling and simulating complex systems {{cite:6888568513677175a6f3f92abacd11b880ca615a}}. Also, it has grown enormously not only applied to traffic but also to transportation in general terms {{cite:4e2045e05d24e18f1e7163595aa67d6761a81950}}.",
"In the last decade some microscopic simulators have been developed, such as MITSIM, Paramics, Aimsun, CORSIM and VisSim. However, none of these is strictly defined as agent-based simulation systems, even though they model vehicles in an object-oriented manner. Just a few simulation tools truly support the concept of agents and multi-agent systems in traffic simulation; MATSim-T {{cite:35f5e3d11756e7bd044eb04a837c10b35f9c3c5d}}, {{cite:7ad9665af3cdcbf18531e765faabdbd63226f8e4}} and ITSUMO {{cite:8b925c3be0530c32ea60e36c96374b3b507c2ed6}}, {{cite:5a9bf0a094a821af26ae5ff08c87610f3a699b68}} are good examples to be mentioned.",
"Regarding this simulation tools some examples of multi-agent system approaches for traffic lights control can be seen in {{cite:0dc144695db823ae9258cd929a10a1bbcb328e6c}}, {{cite:03622414762ce154962d7ccd1efc89977aeb101a}}, {{cite:f4cb3b6ae26b543961887b500b9efc7e35c5e1df}} and {{cite:aeb4c643d103adedf6573a09385d2423a4f5bd37}}. Simulators used in these works were Aimsun, ITSUMO, VisSim and ITSUMO, respectively.",
"With MAS being recognized as a convenient approach, there must be a sufficiently general way to couple this approach to such a huge quantity of microscopic simulators that exist now. The platform that integrates SUMO and JADE consists of an API intended to allow interoperability among simulators. The platform, coined TraSMAPI, is sufficiently general to allow other simulators to interact with MAS frameworks such JADE. A previous paper {{cite:4347fc76255ae09c4341247e61ea74c3c4b23fcb}} reports on an experiment integrating ITSUMO and SUMO under TraSMAPI, thus demonstrating such an ability. In another study {{cite:b9d11acdb83d2353a57116fbfd3f559b6e0ddb53}}, external traffic controller agents operate over Aimsun-simulated scenarios through TraSMAPI. In this specific work, we illustrate how non-agent-based simulators can be extended with TraSMAPI to support MAS-T assessment. There are certainly other options to simulate agent-based traffic and transportation, such as MATSim. Although such tools are open-source then allowing full customisation, the use of JADE over a traditional microscopic simulation tool is expected to promote greater flexibility in terms of agent architectures that can be implemented.",
"In respect to the described tool-chain, a similar approach has already been proposed. In {{cite:1e755cc5e29818adcb3c0bd4469bf3f30853f8dc}} it is possible to see the tool-chain JADE+TraSMAPI+SUMO. However, the goal of this work was focused on the vehicles itself instead of traffic lights."
] |
[] |
Time Resource Networks
|
The problem of scheduling under resource constraints is widely applicable.One prominent example is power management, in which we have a limitedcontinuous supply of power but must schedule a number of power-consuming tasks.Such problems feature tightly coupled continuous resource constraints andcontinuous temporal constraints. We address such problems by introducing the Time Resource Network (TRN), anencoding for resource-constrained scheduling problems. The definition allowstemporal specifications using a general family of representations derived fromthe Simple Temporal network, including the Simple Temporal Network withUncertainty, and the probabilistic Simple Temporal Network (Fang et al.(2014)). We propose two algorithms for determining the consistency of a TRN: one basedon Mixed Integer Programing and the other one based on Constraint Programming,which we evaluate on scheduling problems with Simple Temporal Constraints andProbabilistic Temporal Constraints.
|
[
"One of the earliest mentions of a scheduling problem being solved in an algorithmic fashion can be found in {{cite:38bfbb2ddbcd371a0fbb56c9914439195f0cd966}}, although there's evidence that the problem was already considered in unpublished versions of {{cite:f84827a3f002274f01a526873395273dbaf66d16}}. This publication considers the following statement of scheduling problem. We have {{formula:4a7a0af3-c6c8-49ee-a1f1-590b3fad9ff6}} items and {{formula:3f218701-42e2-4b44-b7f1-987829716c45}} stages and {{formula:147db5b0-da32-4212-8970-a9644a5fcd8c}} denoting the time for {{formula:93e19008-c434-4c39-8891-3eed82cba88e}} -th item to be processed by stage {{formula:a3566dc6-4ba7-408a-b57e-420bdd8d64de}} . All the items must be processed by different stages in order (for example first stage is printing of a book and second stage is binding). The publication considers {{formula:4f6144a3-8f3a-4b13-b6ee-d3775db38cd0}} and {{formula:027ca60b-1c09-4a39-8012-9f257c93dca0}} and arrives at the solution that “permits one to optimally arrange twenty production items in about five minutes by visual inspection”. It turns out that the solution to the problem for {{formula:21b32b06-6602-44e8-be92-36b8bd4701c8}} is NP-hard ({{cite:3404612c11b6e5ebce8a63a50c1c8b01f016fc1d}}). In {{cite:80e1269759f39b8887ff658545b89b86dbe2c488}} an Integer Programming solution to the scheduling problem was presented, with a comment stating that it “is a single model which encompasses a wide variety of machine-scheduling situations”.",
"In {{cite:1dda9358758ef0c79c61fbae0bb8af86c548f61b}}, a generalization of scheduling problems is considered, which allows for multiple resource constraints. However, the proposed solution uses a discrete time formulation, which, depending on required accuracy, can substantially decrease performance. In 1988 a technique was proposed which can handle resource constraints and continuous time ({{cite:74e38f1fdfd44e12a180a3f9739723d65b7b2b58}}). The proposed approach can be thought of as resource constrained scheduling over Simple Temporal Networks (STN).",
"In {{cite:5bfc9da15eceef3fb5cd8e82f961b51f9d572408}}, a notion of Simple Temporal Problem was introduced which allows one to solve problems with simple temporal constraints of form {{formula:7f4ba405-ddcd-432e-a25f-67a62093b0dd}} . This concept was later extended with various more sophisticated notions of temporal constraints. {{cite:75a8e4bcbb745d04e58e54d1f4059e6066dbe60b}} defined an uncertain temporal constraint, where the duration between two time events can take a value from an interval {{formula:c1021898-c9f3-49eb-bede-ae175d377fb5}} , which is unknown during the time of scheduling (uncertain duration constraints). {{cite:edf6279f7b2f7463314782a8b9139d87d355f4d9}} describes a pseudopolynomial algorithm for handling uncertain duration constraint, where we are allowed to make a scheduling decisions based on knowledge of uncertain durations from the past (Dynamic controllability). The algorithm is later improved to polynomial complexity ({{cite:c977df78ef233efd97e189bec60a6ec57e21193b}}). Finally, {{cite:ecc90cc3457633668a9186528d4318068bb05efa}} provides a non-linear optimization based solver for uncertain temporal constraints where the duration of the constraint can come from arbitrary probabilistic distribution."
] |
[] |
DAP3D-Net: Where, What and How Actions Occur in Videos?
|
Action parsing in videos with complex scenes is an interesting butchallenging task in computer vision. In this paper, we propose a generic 3Dconvolutional neural network in a multi-task learning manner for effective DeepAction Parsing (DAP3D-Net) in videos. Particularly, in the training phase,action localization, classification and attributes learning can be jointlyoptimized on our appearancemotion data via DAP3D-Net. For an upcoming testvideo, we can describe each individual action in the video simultaneously as:Where the action occurs, What the action is and How the action is performed. Towell demonstrate the effectiveness of the proposed DAP3D-Net, we alsocontribute a new Numerous-category Aligned Synthetic Action dataset, i.e.,NASA, which consists of 200; 000 action clips of more than 300 categories andwith 33 pre-defined action attributes in two hierarchical levels (i.e.,low-level attributes of basic body part movements and high-level attributesrelated to action motion). We learn DAP3D-Net using the NASA dataset and thenevaluate it on our collected Human Action Understanding (HAU) dataset.Experimental results show that our approach can accurately localize, categorizeand describe multiple actions in realistic videos.
|
[
"Since little work has been done on action parsing in videos for simultaneously solving the problems of action localization, categorization and attributes learning, in this section, we mainly review some related work on action detection and action attributes modeling. Action detection can be regarded as a combination of action localization and categorization. In {{cite:05d82d0ebd9d8fb036551216bddffca38fa84112}}, a weakly supervised model with multiple instance learning was applied for action detection. In {{cite:6c1596b0449a4a3250cabb5c1cf8e98facdda921}}, a dynamic-poselets method was introduced. Branch-and-bound algorithm {{cite:3b9772d391011a26ef451d7167440849e0ffa101}} was proposed to reduce the action detection complexity. There also exist some sub-volume {{cite:a5841332e0f8fef3d762adf901cd245b2bd31546}}, {{cite:f1857fee2cf57e91c377b1e8881b969a1a55123a}}, {{cite:c04076fc8a50d81ec0c1f23ac9bd67e9c2c41b02}} based action detection methods. Besides, the cross-dataset action detection {{cite:404cdcd175003d8c4e29add9af97723dc5434bb7}} and the spatio-temporal deformable part models based action detection {{cite:57d44ee213f3d815ede88eda48fad8855abbfe50}} have also been proposed in previous studies. Additionally, action detection via fast proposals was developed in {{cite:66db122f39c712b6a4eb1c84808183e61c902528}}.",
"For action attributes modeling, Liu et al. {{cite:f802c5e399af8051cfb9d744dc6edb31f2c3a299}} used high-level semantic attributes to represent human actions in videos and further constructed more descriptive models for the action recognition task. A similar idea has also been applied in {{cite:ed07de757de960130f1166baaadb432a2b4792fd}}, {{cite:31ec019289f98614fefd589035c0e254231e837a}} for improved action categorization. Moreover, a convolutional multi-task learning method {{cite:5c754c1fbfd5b9a980ade664ca4abcd69b8b8668}} has been adopted for action recognition from low-level features with attribute regularization. In {{cite:911e239ca70cbd622cc383c84af15f81fd527656}}, a robust learning framework using relative attributes was developed for human action recognition. Additionally, action attributes and object-parts from images were also used for action recognition in {{cite:890e06ba212e9235cf0d1a2c7f3410e21f3e5be0}}. However, all the above studies mainly focus on action recognition by means of attributes rather than general action attributes learning tasks. Although, in {{cite:e73b0336e97828cc2b6da4db9f03a518f18b11c4}}, authors have jointly tackled the classification and attributes annotation for group activities, it is still regarded as a separated feature extraction and attribute learning pipeline rather than an end-to-end framework as our multi-task DAP3D-Net. Besides, DAP3D-Net focuses on simultaneous localization and parsing of multiple actions, while in {{cite:e73b0336e97828cc2b6da4db9f03a518f18b11c4}} only global representations of group activities are considered."
] |
[] |
Package equivalence in complex software network
|
The public package registry npm is one of the biggest software registry. Withits 216 911 software packages, it forms a big network of software dependencies.In this paper we evaluate various methods for finding similar packages in thenpm network, using only the structure of the graph. Namely, we want to find away of categorizing similar packages, which would be useful for recommendationsystems. This size enables us to compute meaningful results, as it softened theparticularities of the graph. Npm is also quite famous as it is the defaultpackage repository of Node.js. We believe that it will make our resultsinteresting for more people than a less used package repository. This makes ita good subject of analysis of software networks.
|
[
"The modern science of network is particularly interested in decomposing nodes of large networks into independent groups called \"communities\" {{cite:a9cbc1de54c6a34f62ce38662d69b8963188411f}}. As a community, you think of a group of nodes that is internally densely connected but sparsely connected externally {{cite:3be3eb72857bc9e7ab1c1074f9207644be473f66}}, {{cite:ac208bf2cf809a40d9ddbaa54f7f139c8a614f9c}}, {{cite:a9cbc1de54c6a34f62ce38662d69b8963188411f}}, {{cite:464df4b36e7e112542223b4936e68efa0c8e5e7f}}. The nodes within community are similar to each other and dissimilar to the rest of the nodes in a network. Researchers showed there is actually community structures in real-world networks {{cite:34b05b658094374d60893d3e6f8b0295812f1c79}}, {{cite:a9cbc1de54c6a34f62ce38662d69b8963188411f}}, {{cite:3c2dae901b7a7454c1d888325be7e240e7e32fda}}. However, the definition of community is not universally accepted and for that matter we have multiple definitions. Community structure emphasizes cohesive groups of nodes and the absence of dependencies between the groups, but this does not say anything about the roles in a network. The concept of roles in networks is much wider that the concept of community. Prerequisite for the analysis of roles in a network is a community structure. After that, you examine how discovered communities are inter-dependent, which translates to different roles in a network.",
"One of predominant technique for deriving structure of a network is blockmodeling {{cite:93f182035ac3f12050414ed59a388c7cca1dc9c5}}. It is considered as mapping nodes and edges onto their images in a reduced graph. A node (\"block\") in reduced graph represents the nodes from original graph that were mapped to the same image. This type of mapping is called semigroup homomorphism {{cite:4e8ebd811334e1850a24215bae2aafb9e24a332a}}. Homomorphisms are mappings that preserve the structure of a graph in a way that nodes of a graph are mapped into nodes in an image of the graph, and each edge in an image of the graph is connected to the same nodes as in the original graph. The semigroup is an algebraic structure consisting of a set with an associative binary operation. We can look at the nodes from original graph as a semigroup. Consequently, the nodes in reduced graph represent the result of applying the semigroup operation to the ordered pair {{formula:ce154ffb-9ff2-49e8-8d74-1639b45df1d0}} , which in our case are the edges in the original graph. If we take equivalence relation as the operation within semigroup, then we get equivalence classes as the nodes in the reduced graph. The plot of adjacency matrix of the new graph reveals blocks in diagonal, which are equivalence classes, and dependencies between them.",
"There exists a framework for blockmodeling classes within complex networks {{cite:3be3eb72857bc9e7ab1c1074f9207644be473f66}}. Authors derive a measure to find the best fitting image graph (reduced graph) and present a criterion to avoid overfitting. Image graph is termed as the role model. They did not demand exact mapping of every single node to the role model, but that the network as a whole fit as well as possible to the role model. The perfect fit would correspond to regular and structural equivalence. When tackling a new network, they assumed a given image graph and assignment of roles to nodes. They derived a quality function as an objective measure of fit between the image and the network under this assignment of roles. The assignment of roles which maximizes their quality function is considered as the best one to describe the connection structure of the original network. The concepts of modularity {{cite:ef2fc7cd99b65b19d3586c6d39f85054bc3b4a68}} and structural equivalence are found as special cases of this approach. The modularity measure is commonly used in community detection algorithms. It measures the quality of a network division into communities.The proposed method is applicable to both two-mode and one-mode data, directed and undirected, as well as weighted networks. It is non-parametric and computationally efficient. In the same paper authors applied method to the world trade network and analyzed the roles individual countries play in the global economy.",
"Another research dealt with analysis of patterns of role-to-role connections, but with their own definition of roles {{cite:64abe28ce253195a1f0f42a9e1fd74c5582d2f4a}}. Principle was the same; to group the nodes into roles, according to their pattern of intra- and intergroup connections. They analyzed four different types of real-world networks; metabolic networks, protein interactions, global and regional air transportation networks, and the Internet at the level of autonomous systems. To determine and quantify the modular structure of these networks, they use simulated annealing. The optimal partitions of the network into groups was found with the use of the modularity measure. By comparing modular structure of each network with the randomization of the same network, they found out that all observed networks have a significant modular structure. That is reasonable as groups in biological networks corresponds to functional units and in air transportation groups corresponds to geo-political units. The role of each node was determined according to two properties: the relative within-group degree {{formula:53dcecee-4d4e-4716-b5d7-28c0ec506771}} , which measures the node's degree of connectedness with the nodes within the same group, and the participation coefficient {{formula:fc104e20-c5e1-4a15-9bae-8e5386f953f7}} , which measures node's degree of connectedness with other groups. When classifying nodes into roles, they initially measured within-group degree of all nodes, and divided them into hubs and non-hubs based on high and low within-module degree, respectively. Then they looked at the participation coefficient to further subdivide hubs into different types of hubs. Further subdivision was also done to non-hubs. Reasoning behind this particular definition of the roles is given in {{cite:3c2dae901b7a7454c1d888325be7e240e7e32fda}}."
] |
[] |
Reinforcement Learning approach for Real Time Strategy Games Battle city
and S3
|
In this paper we proposed reinforcement learning algorithms with thegeneralized reward function. In our proposed method we use Q-learning and SARSAalgorithms with generalised reward function to train the reinforcement learningagent. We evaluated the performance of our proposed algorithms on two real-timestrategy games called BattleCity and S3. There are two main advantages ofhaving such an approach as compared to other works in RTS. (1) We can ignorethe concept of a simulator which is often game specific and is usually hardcoded in any type of RTS games (2) our system can learn from interaction withany opponents and quickly change the strategy according to the opponents and donot need any human traces as used in previous works. Keywords : Reinforcementlearning, Machine learning, Real time strategy, Artificial intelligence.
|
[
"One of the major works using Online case-based planning {{cite:e61787633f867a4e1aac64e51659ef9e3681db7a}} techniques for Real Time Strategy Games was published in {{cite:dae1d3f41f6dca3c74facc5ffa3d8521be473f2a}}. On-line case-based planning revises case based planning for strategic real-time domains involving on-line planning.",
"In {{cite:82b0ee5312e8aac7b9c62e7c0e736202d00984e4}} a case-based planning system called Darmok2 is introduced that can play RTS games. They introduced a set of algorithms that can be used to learn plans, represented as petri-nets, from one or more human demonstrations.Another work by the same authors which uses Darmok2 but addresses the issues of plan acquisition, on-line plan execution, interleaved planning and execution and on-line plan adaptation is {{cite:bedd514b32bf40ea030724cc682324611c9b6a30}}.",
"In {{cite:fc62f3a8246cab6b32cb0e7da4c7d9b08c6a1c71}} the authors summarize their work in exploring the use of the first order inductive learning (FOIL) algorithm for learning rules which can be used to represent opponent strategies.",
"In {{cite:b4a25406f48712c603ea3856274aa0e7364939fd}} the authors improve Darmok2 using information related to sensors of the game. We refer to that work as PR-Model in this paper. PR-model is capable of learning how to play RTS games by observing human demonstrations. Using human traces PR-model makes plans to play games. Prioritize the plan according to the feedback of the game and feedbacks are decided using some rule which depends on the sensors of the game.",
"Drawbacks of all case based learning {{cite:c0e227b91db250fb4529f27e583ebc5a03007366}} approaches as mentioned above are (1) It requires expert demonstrations for making plans (2) after training is done, no further learning takes place (3) to cover large state spaces it would require large number of rules in the plan base (4) no exploration for optimal solution. Only follows human traces.. Stefan Wender {{cite:55c5f9cdaba1edce0f789b37dfb4124e6c79a294}} uses Reinforcement Learning for City Site Selection in the Turn-Based Strategy Game Civilization IV. Civilization IV is the strategy game it is a turn-based game while Battle City is Real time game.",
"Stefan Wender {{cite:55c5f9cdaba1edce0f789b37dfb4124e6c79a294}} uses Reinforcement Learning for City Site Selection in the Turn-Based Strategy Game Civilization IV. Civilization IV is the strategy game similar to S3 but it is a turn-based game while S3 is Real time multi agent game.",
"In this paper we aim to do away with the hard coded simulator and propose a learning approach based on Reinforcement Learning {{cite:260123f21de5e09d72a8a29866f7dd1c5f2ec29c}}(RL) wherein sensor information from the current game-state is used to select the best action.Reinforcement learning is used because of its advantages over previous strategies. Specifically (1) RL cuts out the need to manually specify rules. RL agents learn simply by playing the game against other human players or even other RL agents (2) for large state spaces, RL can be combined with a function approximator such as a neural network, to approximate the evaluation function (3) RL agent always explores for optimal solution to reach the goal (4) RL has been applied widely to many other fields, such as robotics, board games ,turn based games and single agent games with great results, but hardly ever on RTS multi-agent games."
] |
[] |
Spectral Learning for Supervised Topic Models
|
Supervised topic models simultaneously model the latent topic structure oflarge collections of documents and a response variable associated with eachdocument. Existing inference methods are based on variational approximation orMonte Carlo sampling, which often suffers from the local minimum defect.Spectral methods have been applied to learn unsupervised topic models, such aslatent Dirichlet allocation (LDA), with provable guarantees. This paperinvestigates the possibility of applying spectral methods to recover theparameters of supervised LDA (sLDA). We first present a two-stage spectralmethod, which recovers the parameters of LDA followed by a power update methodto recover the regression model parameters. Then, we further present asingle-phase spectral algorithm to jointly recover the topic distributionmatrix as well as the regression weights. Our spectral algorithms are provablycorrect and computationally efficient. We prove a sample complexity bound foreach algorithm and subsequently derive a sufficient condition for theidentifiability of sLDA. Thorough experiments on synthetic and real-worlddatasets verify the theory and demonstrate the practical effectiveness of thespectral algorithms. In fact, our results on a large-scale review ratingdataset demonstrate that our single-phase spectral algorithm alone getscomparable or even better performance than state-of-the-art methods, whileprevious work on spectral methods has rarely reported such promisingperformance.
|
[
"Based on different principles, there are various methods to learn supervised topic models. The most natural one is maximum-likelihood estimation (MLE). However, thehighly non-convex property of the learning objective makes the optimization problem very hard. In the original paper {{cite:f1839e6fb69a664c5a9a64a296da58f6453b0561}}, where the MLE is used, the authors choose variationalapproximation to handle intractable posterior expectations. Such a method tries to maximize a lower bound which is built on some variational distributions and a mean-field assumption is usually imposed for tractability. Although the method works well in practice, we do not have any guarantee that the distribution we learned is close to the true one. Under Bayesian framework, Gibbs sampling is an attractive method which enjoys the property that the stationary distribution of the chain is the target posterior distribution. However, this does not mean that we can really get accurate samples from posterior distribution in practice. The slow mixing rate often makes the sampler trapped in a local minimum which is far from thetrue distribution if we only run a finite number of iterations.",
"Max-margin learning is another principle on learning supervised topic models, with maximum entropy discrimination LDA (MedLDA) {{cite:4cfc09590ae0be9cde1bd2dbcf6835567c46e8bc}} as a popular example. MedLDA explores the max-margin principle to learn sparse and discriminative topic representations. The learning problem can be defined under the regularized Bayesian inference (RegBayes) {{cite:671e2fab77d7e1a165d565c4e53d399f0f0033ee}} framework, where the max-margin posterior regularization is introduced to ensure that the topic representations are good at predicting response variables.Though both carefully designed variational inference {{cite:4cfc09590ae0be9cde1bd2dbcf6835567c46e8bc}} and Gibbs sampling methods {{cite:746a39be361b183b51cc50b05334e6010b6a57e0}} are given, we still cannot guarantee the quality of the learnt model in general.",
"Recently, increasing efforts have been made to recover the parameters directly with provable correctness for topic models, with the main focus on unsupervised models such as LDA.Such methods adopt either NMF or spectral decomposition approaches.For NMF, the basic idea is that even NMF is an NP-hard problem in general, the topic distribution matrix can be recovered undersome separable condition, e.g., each topic has at least one anchor word.Precisely, for each topic, the method first finds an anchor word that has non-zero probability only in that topic. Then a recovery step reconstructs the topic distribution given suchanchor words and a second-order moment matrix of word-word co-occurrence {{cite:623e346763109dd1681d766440a8f5fe1a3f9bdc}}, {{cite:33f364ccf74f509bd11bbc4fa3f25aa96933a91b}}.The original reconstruction step only needs a part of the matrix which is not robust in practice.Thus in {{cite:33f364ccf74f509bd11bbc4fa3f25aa96933a91b}}, the author recovers the topic distribution based on a probabilistic framework.The NMF methods produce good empirical results on real-world data.Recently, the work {{cite:065d1d8b84ab4c15c00750bf6d89072d9267bac0}} extends the anchor word methods to handle supervised topic models. The methodaugments the word co-occurrence matrix with additional dimensions for metadata such as sentiment, and shows better performance in sentiment classification.",
"Spectral methods start from computing some low-order moments based on the samples and then relate them with the model parameters.For LDA, tensors up to three order are sufficient to recover its parameters {{cite:5ab384f2ca55ae19a0dcafea77937e2e144d6e28}}.After centralization, the moments can beexpressed as a mixture of the parameters we are interested in.After that whitening and robust tensor decomposition steps are adoptedto recover model parameters. The whitening step makes that the third-order tensor can be decomposed as a set oforthogonal eigenvectors and their corresponding eigenvalues after some operations based on its outputand the robust tensor decomposition step then finds them.Previous work only focuses on unsupervised LDA models andwe aim to extend the ability for spectral methods to handle response variables. Finally,some preliminary results of the two-stage recovery algorithm have been reported in {{cite:7fabee642daa5170c524e73073027e48f0225412}}.This paper presents a systematical analysis with a novel one-stage spectral method,which yields promising results on a large-scale dataset."
] |
[] |
Improving Trajectory Modelling for DNN-based Speech Synthesis by using
Stacked Bottleneck Features and Minimum Generation Error Training
|
We propose two novel techniques --- stacking bottleneck features and minimumgeneration error training criterion --- to improve the performance of deepneural network (DNN)-based speech synthesis. The techniques address the relatedissues of frame-by-frame independence and ignorance of the relationship betweenstatic and dynamic features, within current typical DNN-based synthesisframeworks. Stacking bottleneck features, which are an acoustically--informedlinguistic representation, provides an efficient way to include more detailedlinguistic context at the input. The minimum generation error trainingcriterion minimises overall output trajectory error across an utterance, ratherthan minimising the error per frame independently, and thus takes into accountthe interaction between static and dynamic features. The two techniques can beeasily combined to further improve performance. We present both objective andsubjective results that demonstrate the effectiveness of the proposedtechniques. The subjective results show that combining the two techniques leadsto significantly more natural synthetic speech than from conventional DNN orlong short-term memory (LSTM) recurrent neural network (RNN) systems.
|
[
"Very substantial effort has been devoted to acoustic modelling in the hidden Markov model (HMM) speech synthesis framework. Amongst the many proposed techniques, we highlight just a few of the most influential. In {{cite:c795108f6516d2a046fb39174d73061f6cce38f8}}, a minimum generation error training criterion was proposed to address an inconsistency between training and generation criteria, and the lack of interaction between static and dynamic features during training. In {{cite:a24cfa45e698ab9b4cc90aa46940d189f5d77f67}}, the so-called trajectory HMM was proposed to explicitly model the relationships between static and dynamic features. As a complement to improving the acoustic model itself, enhancement techniques such as global variance {{cite:93af096afe71feab69102276dd4d5daf18707bd4}} and modulation spectrum enhancement {{cite:619a55c96a680252b7a8a43cbbee06cd970f0a9f}} aim to mitigate the lack of variation in generated parameter trajectories that results from using an incorrect acoustic model. Although such enhancement techniques do not reduce objective error (e.g., lower spectral distortion w.r.t. a natural speech reference), significant improvements in subjective naturalness are obtained. However, none of the above techniques address what is perhaps the most fundamental problem of HMM-based speech synthesis: across-context averaging via decision tree clustering, which has been identified as a major contributing factor to reduced naturalness {{cite:d16cee9ff3e03df527dd07265a17b37e4286cd7a}}.",
"More recently, following on from successes in automatic speech recognition {{cite:980c5ff81c7601132ca24d3fe035b2c04156d987}}, artificial neural networks have re-emerged as acoustic models for SPSS {{cite:cf5103d5ba3e0bd394e3736082222e7c92be8d0d}}.By the 1990s, artificial neural networks had already been employed as feature extractors from text input to produce linguistic features {{cite:841a50f596b2e1a4670c15aa6db33a2c870bbccf}}, as acoustic models to map linguistic features to vocoder parameters {{cite:a87d524af0ba6aea08bb81d4659545a8ffe85448}}, {{cite:9ccb4c6c383876bfddb4ebb6a51e14a5568a8969}}, {{cite:152cb9d4ff0c4632c0b28f9e7024e4ac53f5c169}}, and to predict segment durations {{cite:f525f7e02194fab8fb535f0d3e4e67fe93952428}}. One prominent theme in more recent studies is the use of neural architectures to replace Gaussian mixture models (GMMs) associated with leaf nodes of decision trees, such as the restricted Boltzmann machines (RBMs) in {{cite:0e041566490714981917bab4cdb5b6adf3aa18d9}}, where RBMs were claimed to better learn spectral detail, resulting in better quality synthesised speech. In {{cite:89d7ec7e4f142647400541e139464c04bd3d767f}}, {{cite:5b74027cffe9b82ec4d047f257ba102ba25b6bd7}}, a deep belief network (DBN) was employed as a deep generative model of the joint probability distribution between linguistic and acoustic features. Other variants on the neural architectures applied to SPSS include the use of deep mixture density networks to predict probability density functions over acoustic features given the corresponding linguistic features {{cite:aa79446eb6e592a0a000043290a71788000ad159}} and a trajectory real-valued neural autoregressive density estimator to model acoustic parameter trajectories as well as across-feature dependencies {{cite:f7684750103e8c0f7840b8f791a368c242514275}}. Neural approaches have been applied to enhancement too, such as the deep generative model in {{cite:d9435a3517f0a970a29700a74b8bc22829b5991b}} acting as a post-filter to enhance the quality of speech synthesised from an HMM-based system.",
"The most popular way to use neural networks in SPSS is with a deep feed-forward neural network (DNN) as a conditional model to map linguistic features to vocoder parameters directly {{cite:4887075ee2b8052b44ba43887a809c5ff5375e54}}, {{cite:5988b0537db01bd4c259d4eba00f74848f3b2ec1}}, {{cite:150b79ae1019eadae93ed0f6306556ec9a127e71}}, {{cite:6142260f29e4aae3c4ae24387916a78a82769732}}, {{cite:034a56ded1ba5fcc49dca53c5974ecdf03f289b5}}. This can be viewed as replacing the decision tree used in HMM-based speech synthesis with a more powerful regression model {{cite:4887075ee2b8052b44ba43887a809c5ff5375e54}}, {{cite:0dc5882fecfcfae99546da42f5c9346986b3b91c}}. DNNs have other advantages, including the ability to model high-dimensional acoustic parameters (e.g., the spectrum {{cite:92c9c219c0b2bfb6633b18cc05f401a18a6badf7}}), and the availability of techniques such as multi-task learning {{cite:6142260f29e4aae3c4ae24387916a78a82769732}}, {{cite:ebef2c95c5a8b68ce919fcce446bee92a3cff371}}. However, a limitation of standard DNN implementations is that the mapping is performed frame by frame without considering contextual constraints, other than those encoded in the input linguistic features. Even though dynamic features are part of the output acoustic feature vector – because they are needed as a constraint to generate smooth parameter trajectories at synthesis time – contextual constraints between statics and deltas (or equivalently between successive frames of statics) are not explicitly modelled during training.",
"One way to model contextual constraints is proposed in {{cite:b82b200e866338c83caad05d9f29063ef4d27dfc}}: a bidirectional long short-term memory (LSTM)-based recurrent neural network (RNN) to map a sequence of linguistic features to the corresponding sequence of acoustic features. An LSTM with a recurrent output layer is proposed in {{cite:a792aedf353b78b56382d170ca575dcba53dbbaa}} to smooth acoustic features across consecutive frames. A systematic investigation on the architectures of gated recurrent neural network can be found in {{cite:ffc517a8bd468d227b0387846c41db27a2390c11}}. These studies formulate speech synthesis as a sequence-to-sequence mapping problem and provide evidence that a better model of speech parameter trajectories results in better synthetic speech."
] |
[] |
Automatic Generation of Formula Simplifiers based on Conditional Rewrite
Rules
|
This paper addresses the problem of creating simplifiers for logic formulasbased on conditional term rewriting. In particular, the paper focuses on aprogram synthesis application where formula simplifications have been shown tohave a significant impact. We show that by combining machine learningtechniques with constraint-based synthesis, it is possible to synthesize aformula simplifier fully automatically from a corpus of representativeproblems, making it possible to create formula simplifiers tailored to specificproblem domains. We demonstrate the benefits of our approach for synthesisbenchmarks from the SyGuS competition and automated grading.
|
[
"A pre-processing step in constraint solvers and solver-based tools (like Z3, Boolector {{cite:5e6b141619b70865bcf4512b8e6dc63d7019d5ac}}, Sketch etc) is an essential one and term rewriting has been extensively used as a part this pre-processing step {{cite:bce290312bc7b2b84dbb09f7a523e5ce0dae27b4}}, {{cite:3bbca0d12558334ed0bdd4a51bfd33afb1b199e3}}, {{cite:3ac4de6832750fbb5274867bd0cccb34f46921df}}, {{cite:65f61079a9271d41e3ccf6d4b24a8671d970a77f}}, {{cite:6f5dd54bd506bf2c1c4372db91f1deaadfd36a4f}}, {{cite:c710784444938b57b07c74f4c0ad4c9498e75733}}. These pre-processing steps are very important and can have a significant impact on performance. As others have observed and our experiments confirm,however, rewrites that help for one class of problems may not be effectivefor problems for a different domain. The task of developing domain-specific simplification rules is time consuming and error-prone, and our technique aims to automate the discovery of such rules.",
"Each part of our framework solves an independent problem and is different from the state of the art, specialized for our purposes. A recent paper introducing Alive {{cite:4bb6ddfcf29ee10290afc7052da3214a55e4cd96}}, a domain specific language for specifying, verifying, and compiling peephole optimizations in LLVM is the closest to our framework as a whole. Their rewrite rules are guarded by a predicate, they use static analyses to find the validity of those guards, they verify the rules and then compile them to efficient C++ code for rewriting LLVM code: all similar to our phases. However, their system is targeted towards the compilers community and relies upon the developers to discover and specify rewrite rules. Our work is targeted towards the solver community and automatically synthesizes the rewrite rules from benchmark problems of a given domain.",
"Identification of recurrent sub-graphs from benchmark DAGs is similar in essence to the Motif discovery problem {{cite:67b14da89f06090a546f6d852500bf28c7302e0b}} which is famous because of its application in DNA fingerprinting{{cite:53734a07973f78cf1aba7716793b3f783b7dd336}}. This is a very active area of research and recently we have seen some attempts to use sampling {{cite:e6f9281df5425137ee28596142d2374e44aa7f07}}, machine learning {{cite:44bec5c5dbabacbd3f616594592f795f4a80b888}} and distributed algorithms {{cite:eaa3416a54d616c4b3fc4f7cba7749d4c25c15e3}}, {{cite:e9e4e4572a6e7c16210ff0cc83f6001d3112d2b7}}, {{cite:81ca72ba796820195efc3f43f83522c0eec209cd}} to compute the Motifs (statistically significant recurrent sub-graphs) as quickly as possible. Our DAGs, on the other hand, have labeled nodes (labeled with operation types) and our motifs have to account for the fact that some opperations are commutative but not others, which makes direct translation to Motif discovery problem more difficult.",
"In the superoptimization community, people explore all possible equivalent programs and find the most optimal one. It would not make sense to do that for formulas. But {{cite:f09f63b5a5cd9594786aa7074e69394ecc4485d5}} came up with this idea of packaging the superoptimization into multiple rewrite rules similar to what we are doing here. Although it looks similar in spirit to our work, there are a few differences. Most importantly, {{cite:f09f63b5a5cd9594786aa7074e69394ecc4485d5}} uses enumeration of potential candidates for optimized versions of instruction sequences and then checks if it is indeed the most optimal version. Whereas, we use a hybrid approach that primarily relies on constraint based synthesis for generating the rules, which offers a possibility of specifying a structured grammar for the functions.",
"The third phase in Swapper automatically generates simplifier's code (representing an abstract reduction system) is similar to a term or graph rewrite system like Stratego/XT {{cite:452ccfba37eac3a3a97e9c46996847702a6e7f91}}, {{cite:5aec4ee32c456eb59264a4ee18b9f06892e691bf}} or GrGEN.NET {{cite:abf1bebdd72b6ebbfc73470fa3660f99df451ab9}}. Stratego/XT is a framework for the development of program transformation systems and GrGEN.NET is a Graph Rewrite Generator. They offer declarative languages for graph modeling, pattern matching, and rewriting. Both of these tools generate efficient code for program/graph transformation based on rule control logic provided by the user.We build upon their ideas and develop our own compiler because we already had special input graphs (DAGs with operations as labels of nodes) and an existing framework for simplification (Sketch's DAG simplification class). Our strategy can be compared with LALR parser generation {{cite:df7431fc2970970255c35b0134e0cf258688ed36}} where the next look-ahead symbol helps decide which rule to use. In our case as well we keep around a set of rules that are potentially applicable based on what the algorithm has seen."
] |
[] |
Auto-JacoBin: Auto-encoder Jacobian Binary Hashing
|
Binary codes can be used to speed up nearest neighbor search tasks in largescale data sets as they are efficient for both storage and retrieval. In thispaper, we propose a robust auto-encoder model that preserves the geometricrelationships of high-dimensional data sets in Hamming space. This is done byconsidering a noise-removing function in a region surrounding the manifoldwhere the training data points lie. This function is defined with the propertythat it projects the data points near the manifold into the manifold wisely,and we approximate this function by its first order approximation. Experimentalresults show that the proposed method achieves better than state-of-the-artresults on three large scale high dimensional data sets.
|
[
"Recent efforts for dealing with noisy data are the denoising auto-encoders (DAEs) {{cite:0dab71a019dee6b98a62c01f541b7ae1fc94e99b}} and contractive auto-encoders (CAEs) {{cite:9adbe1b167bc85fa0fea54e25105210ea1560d21}}. The main starting points of these algorithms are that the learned features in the hidden layer keep the important intrinsic structure of the original data set while unimportant information, such as the noise, is discarded as much as possible. In DAEs, noise is manually added to the training data points; since we know the correspondence between the noisy and clean versions, a constraint is introduced to minimize the gap between these two versions according to some loss function. In CAEs, the Jacobian norm of the function, which maps the input data points to the hidden layer, is minimized for a contractive effect. In the extreme case, CAEs may contract all of the data points in the original space to a single point. This constraint is used to discard noise, but the distance between the input and output points is also considered. In the balance of these two constraints, data on the manifold will be unchanged and data outside the manifold will contract to the manifold.",
"Both DAEs and CAEs focus on training a model such that noisy data are projected onto the manifold and non-noisy data are left alone. When the data points have some noise, i.e., they are distributed around the manifold, the output and hidden layers should keep the important information and at the same time be robust to some degree of noise.This process is done implicitly with both approaches. DAEs introduce artificial noise, and assume that eliminating artificial noise will also help to eliminate real noise, i.e., the ideal noise data point is projected to the original data point through the forward propagation of the learned neural network. For CAEs, the noise discarding process is done through balancing geometric consistency and the contractive property by minimizing the norm of the Jacobian matrix. This motivates us to explore an explicit function which has the ideal noise absorbing property.",
"Fig. REF displays one of the functions we are going to explore. Notice that each data point around the manifold is exactly projected to the nearest point on the manifold, while DAEs and CAEs do not have this kind of guarantee albeit they are designed to discard noise.{{figure:db352450-3c41-4e3f-95fe-5714fb9fb844}}"
] |
[] |
Crowdsourcing On-street Parking Space Detection
|
As the number of vehicles continues to grow, parking spaces are at a premiumin city streets. Additionally, due to the lack of knowledge about streetparking spaces, heuristic circling the blocks not only costs drivers' time andfuel, but also increases city congestion. In the wake of recent trend to buildconvenient, green and energy-efficient smart cities, we rethink commontechniques adopted by high-profile smart parking systems, and present auser-engaged (crowdsourcing) and sonar-based prototype to identify urbanon-street parking spaces. The prototype includes an ultrasonic sensor, a GPSreceiver and associated Arduino micro-controllers. It is mounted on thepassenger side of a car to measure the distance from the vehicle to the nearestroadside obstacle. Multiple road tests are conducted around Wheatley, Oxford togather results and emulate the crowdsourcing approach. By extracting parkedvehicles' features from the collected trace, a supervised learning algorithm isdeveloped to estimate roadside parking occupancy and spot illegal parkingvehicles. A quantity estimation model is derived to calculate the requirednumber of sensing units to cover urban streets. The estimation isquantitatively compared to a fixed sensing solution. The results show that thecrowdsourcing way would need substantially fewer sensors compared to the fixedsensing system.
|
[
"In this section, we overview prominent on-street parking systems and summarize their representative features. By comparing CroPark with these parking systems, the contributions of our work are described.{{table:35548b99-b631-4041-8e09-2e5f011c0bde}}",
"SFpark {{cite:df97eb3f8103bcf137f399275e3be1a33f304cde}} is a U.S. federally-funded parking management project in San Francisco, which spans from 2009 to 2014 with a total budget of {{formula:6cf2fe24-1802-4c99-8bf4-7c9ef95f9cf9}} million dollars. It adopts a classical wireless sensor network structure, where the data flows from parking sensors and parking meters to a data warehouse via wireless mesh networks. The system gives a complete on/off-street parking solution and brings a list of benefits: easier to find parking (searching time decreases by {{formula:51bbab4c-fdff-4b32-a23d-9543a82a56bb}} ), reduced congestion (traffic volume decreases by {{formula:f0a0d11e-be1d-4f64-9228-0ffa9dea221a}} ), lower parking rates (11 cents cheaper), fewer parking tickets ({{formula:e4d31438-3e52-4506-afe7-9c681016f01e}} lower), et cetera. However, the cost of the high-profile parking sysem is rather high. The SFpark pilot deployment installed 11700 magnetometer sensors and 300 pole-mounted mesh nodes for 8000 parking spaces (one or two sensors are installed in each parking space). From SFpark's finance figures, the parking sensors cost {{formula:6541a1a0-85ac-4d9e-911b-8a409fd1c5b2}} million dollars (approximately 480 dollars each), which do not account for management, installation and maintenance expenditure. FASTPRK (magnetic) {{cite:7b21d33fd266903ce9873a1faf0bbdf025b4594a}}, Street Parking System (magnetic) {{cite:c63ce27febaf21fd859ab31ebf1851aa5ca55762}}, Smart Parking (infrared) {{cite:89075e52980e0345c89501b3cde2d17e59f69a62}} and Smart Santander (ferromagnetic) {{cite:f99eac4ae41e985ad9d11b7a3dbe72a6fc2bc18c}} are similar parking projects using different sensing techniques.",
"Integrated Smart Parking Solution {{cite:d21d6e2c369f8b134dfb0c1d712b974ba77ba988}} is a Siemens-led ongoing project aiming at simplifying the searching process, which launched a testing pilot in Berlin in September, 2015. Different from the above sensor-burying approaches, radar sensors are mouted on street lamps to scan a bigger area. The radars not only monitor traffic flows but also parking spaces, which can be either used for traffic control or facilitating drivers to find a parking spot. The benefits of the overhead radar approach are claimed to be as follows. First, it is not impaired by weather or light conditions. Secondly, it detects more than just parking spots. For example, it can measure vehicle speed, traffic flows and pedestrian flows. Thirdly, it is mounted on street lamps, which alleviates the infrastructure changes.{{figure:8f5736fd-db46-40d4-9c1d-8933b35b47fc}}",
"Parking Spotter {{cite:7ca08dc5965adf5b450d0fcb894c29e909e733ff}} is an ongoing joint work by Ford and Georgia Tech. The idea is to leverage sonars and radars that are already widely employed on Ford cars to sense the on-street parking occupancy. The sensing results are updated to a cloud server, and the results are presented to other Ford drivers as an added-value service.",
"Mathur et al. in {{cite:5e491c279f925ffc9147f6c348ec80fcca5e6893}} present ParkNet, a mobile street parking system, which collects parking occupancy information as vehicles pass by. ParkNet consists of a GPS receiver and an ultrasonic rangefinder. Through one-month runs of three vehicles passing by the urban streets of Highland Park, New Jersey, the authors built a parking map from collected data. In order to achieve improved location accuracy, the authors utilize an environmental fingerprinting approach, namely, using objects on the street to correct GPS errors.",
"Nawaz et al. in {{cite:1567c7579f148cfe0f8d5ebb42cf2744d0a5baa6}} propose a WiFi beacon association based sensing system named ParkSense to estimate if a driver has entered or driven away from a parking spot. More specifically, ParkSense uses the WiFi association and de-association changing rates to sense the parking status. From the empirical evaluation, the authors claim that the WiFi beacon rate is highly correlated with driver's activity. The disadvantages of the system are that 1) the presence of both mobile phone and WiFi infrastructure are required, and 2) the access to spatially distributed WiFi access points for analysis is assumed.",
"The features of the surveyed parking systems are summarized in Table REF .",
"The contributions of the article are summarized as follows. Some high-profile parking projects are overviewed. By looking into their features, we rethink the common sensing solutions, and present a crowdsourcing mobile sensing technique. A supervised learning algorithm is developed for CroPark to recognize parked cars and empty spaces. The algorithm uses parked cars' contouring vector extracted from the driving trace to train a classifier, which is used to differentiate parked cars from road clutters. A simple estimation model is proposed to calculate how many mobile sensing units are needed to cover urban streets with certain updating frequency. The estimation produced by the model is quantitatively compared to a fixed sensing solution. The results show that CroPark would need significantly lower number of sensing units than the fixed sensing system."
] |
[] |
Tight Analysis of a Multiple-Swap Heuristic for Budgeted Red-Blue Median
|
Budgeted Red-Blue Median is a generalization of classic $k$-Median in thatthere are two sets of facilities, say $\mathcal{R}$ and $\mathcal{B}$, that canbe used to serve clients located in some metric space. The goal is to open$k_r$ facilities in $\mathcal{R}$ and $k_b$ facilities in $\mathcal{B}$ forsome given bounds $k_r, k_b$ and connect each client to their nearest openfacility in a way that minimizes the total connection cost. We extend work by Hajiaghayi, Khandekar, and Kortsarz [2012] and show that amultiple-swap local search heuristic can be used to obtain a$(5+\epsilon)$-approximation for Budgeted Red-Blue Median for any constant$\epsilon > 0$. This is an improvement over their single swap analysis andbeats the previous best approximation guarantee of 8 by Swamy [2014]. We also present a matching lower bound showing that for every $p \geq 1$,there are instances of Budgeted Red-Blue Median with local optimum solutionsfor the $p$-swap heuristic whose cost is $5 + \Omega\left(\frac{1}{p}\right)$times the optimum solution cost. Thus, our analysis is tight up to the lowerorder terms. In particular, for any $\epsilon > 0$ we show the single-swapheuristic admits local optima whose cost can be as bad as $7-\epsilon$ timesthe optimum solution cost.
|
[
"The study of Budgeted Red-Blue Median from the perspective of approximation algorithmswas initiated by Hajiaghayi, Khandekar, and Kortsarz {{cite:46fa7c1b3ebcccceaa6e1a5e41f1c9a33d74396a}}, where they obtain a constant-factor approximation by a local search algorithm that iteratively tries to swap one redand/or one blue facility in the given solution. They do not specify the constant in their analysis, but it looks to be greater than 8.Citing {{cite:46fa7c1b3ebcccceaa6e1a5e41f1c9a33d74396a}} as inspiration, Krishnaswamy et al. studied a generalization of Budgeted Red-Blue Median known as Matroid Median {{cite:e1fa697fa61ba753c29b8c2a53ead9d4f2e8dfa4}}. Here, a matroid structure is given over the set of facilities and we can onlyopen a set of facilities if they form an independent set in the matroid. They obtain a constant-factor approximation for Matroid Median through rounding an LP relaxation.This was later refined to an 8-approximation by Swamy {{cite:34173779947a06e98a7e1ed41523221d98d843fe}}.",
"The special case of {{formula:476b647f-bd94-4d1a-8ace-fa6985859a43}} -Median is a classic optimization problem and has received a lot of attention from both theoretical and practical communities.The best approximation guarantee known so far is 2.675 by Byrka et al. {{cite:6d0719b830720a0d657f0d33659fe6c44128de0d}}, who build heavily on the breakthrough workof Li and Svensson for the problem {{cite:b6363d71412b2f823396d96992c0017291b73185}}.",
"While local search techniques have been used somewhat infrequently in the design of approximation algorithms in general,it may be fair to say that they have seen the most success in facility location problems.For almost 10 years, the best approximation for {{formula:fd7156f6-3fab-4c4a-8d48-fa1f3b27590c}} -Median was based on a local search algorithm. Arya et al. {{cite:3fb413f29ddffbc6de3ce50a39883ccba28e27f0}} show that a multiple-swap heuristic leadsto a {{formula:deb81bc7-68a2-497b-852f-ca3b82cd9e60}} -approximation for {{formula:e1f3db24-eaf0-4607-8783-1e95a2e0c7e5}} -Median for any constant {{formula:c0679421-8cac-4b08-9df1-08b4b8fb7b41}} . This analysis was simplified in {{cite:9cac1bf79757d8d23aa408b262e8c30ddc128c87}}, which inspires much of our analysis.",
"Other successful local search applications include a {{formula:c91d6121-493e-4a4a-afa0-b9b2d139173b}} -approximation for Uncapacitated Facility Location is also obtained through local search {{cite:3fb413f29ddffbc6de3ce50a39883ccba28e27f0}}, {{cite:a5812447e61d6344ae1a53182cd29cf911166b96}}.Local search has been very helpful in approximating Capacitated Facility Location, the first constant-factor approximation was by Pál, Tardos, and Wexler {{cite:86be33a306e981bf7b5b6e4e931ab226c8c8d932}}and the current best approximation is a {{formula:ca52c177-d791-4404-8e51-1c68dfe95d11}} -approximation by Bansal, Garg, and Gupta {{cite:2b973e6849b0666bc706b3b380f8a9e38a38cea1}}. In the special case whenall capacities are uniform, Aggarwal et al. {{cite:d46c9c04682f548e9bf926c3d728b3c22f41b5c3}} obtain a 3-approximation.Even more examples of local search applied to other facility location variants can be found in {{cite:d921b7b64fa35bd35a4b394af4b784194f6d72e7}}, {{cite:d60bf697737e273616055406b09492f803d95abc}}, {{cite:9cac1bf79757d8d23aa408b262e8c30ddc128c87}}, {{cite:2cb7f4f9abe7d326dbd40fbe6f9f999041af021b}}, {{cite:10f57a77d717730d77955734acc7f4947213032f}}."
] |
[] |
Feedback Control of Real-Time Display Advertising
|
Real-Time Bidding (RTB) is revolutionising display advertising byfacilitating per-impression auctions to buy ad impressions as they are beinggenerated. Being able to use impression-level data, such as user cookies,encourages user behaviour targeting, and hence has significantly improved theeffectiveness of ad campaigns. However, a fundamental drawback of RTB is itsinstability because the bid decision is made per impression and there areenormous fluctuations in campaigns' key performance indicators (KPIs). As such,advertisers face great difficulty in controlling their campaign performanceagainst the associated costs. In this paper, we propose a feedback controlmechanism for RTB which helps advertisers dynamically adjust the bids toeffectively control the KPIs, e.g., the auction winning ratio and the effectivecost per click. We further formulate an optimisation framework to show that theproposed feedback control mechanism also has the ability of optimising campaignperformance. By settling the effective cost per click at an optimal referencevalue, the number of campaign's ad clicks can be maximised with the budgetconstraint. Our empirical study based on real-world data verifies theeffectiveness and robustness of our RTB control system in various situations.The proposed feedback control mechanism has also been deployed on a commercialRTB platform and the online test has shown its success in generatingcontrollable advertising performance.
|
[
"Enabling the impression-level evaluation and bidding, much research work has been done on RTB display advertising, including bidding strategy optimisation {{cite:3ced2f58f4316ac7a4f2a3e27396906f03b372e1}}, {{cite:dfbefc9ed7e90483ec9bcba7e278aa8d9ae54846}}, reserve price optimisation {{cite:6e6ac6807cf726a865baff2a48b70039ff9e47b4}}, ad exchange auction design {{cite:fa33166da7b6cfe73d9083be0240e719e9f656c6}}, and ad tracking {{cite:6d70dd94d703a1d37e23ca215af12eb398ace0ad}}.",
"In order to perform the optimal bidding, the DSP bidding agent should estimate both utility and cost of a given ad impression. The impression-level utility evaluation, including CTR and conversion rate (CVR) estimation, is the essential part for each bidding agent in DSPs. In {{cite:7c3d8b9468e45e0ec4f4bb6a4ab70b1b27098e6c}} the sparsity problem of CVR estimation is handled by modelling the conversions at different hierarchical levels. The user click behaviour on mobile RTB ads is studied in {{cite:f52b31f2c51bcc156a66d0418c570a17353fa478}}. On the cost evaluation side, bid landscape modelling and forecasting is much important to inform the bidding agent about the competitiveness of the market. The authors in {{cite:af7689b3c83fe2a08a45c99cfc03234e61d5802f}} break down the campaign-level bid landscape forecasting problem into “samples” by targeting rules and then employ a mixture model of log-normal distributions to build the campaign-level bid landscape. The authors in {{cite:403fbb1619b4ee6a28ea2f6519d46235dde0eea0}} try to reduce the bid landscape forecasting error through frequently re-building the landscape models. Based on the utility and cost evaluation of the ad inventory, bid optimisation is performed to improve the advertising performance under the campaign budget constraint. Given the estimated CTR/CVR, the authors in {{cite:3ced2f58f4316ac7a4f2a3e27396906f03b372e1}}, {{cite:7c3d8b9468e45e0ec4f4bb6a4ab70b1b27098e6c}} employ linear bidding functions based on truth-telling attributes of second price auctions. However, given the budget constraint, the advertisers' bidding behaviour is not truth-telling. The authors in {{cite:dfbefc9ed7e90483ec9bcba7e278aa8d9ae54846}} propose a general bid optimisation framework to maximise the desired advertising KPI (e.g., total click number) under the budget constraint. Besides the general bid optimisation, the explicit bidding rules such as frequency and recency capping are studied in {{cite:534673d28b7b694a6a9a886ba661006ad0ea14a5}}. Moreover, the budget pacing {{cite:1a7aeb60e717a7a28be008fcae056f9694a2ded2}} which refers to smoothly delivering the campaign budget is another important problem for DSPs.",
"There are a few research papers on recommender systems leveraging feedback controllers for performance maintenance and improvement. In {{cite:8a8bcd1771e26f8da7a9b26df47aa91473e47b6a}}, a rating updating algorithm based on the PID controller is developed to exclude unfair ratings in order to build a robust reputation system. The authors in {{cite:cea4fd1ee06a4a83305b70ebb3e67b612b24a523}} apply a self-monitoring and self-adaptive approach to perform a dynamic update of the training data fed into the recommender system to automatically balance the computational cost and the prediction accuracy. Furthermore, the authors in {{cite:f9ec9d38555f34423a58ebe5f9d0fa234978b5ee}} adopt the more effective and well-studied PID controller to the data-feeding scheme of recommender systems, which is proved to be practically effective in their studied training task.",
"Compared to the work of controlling the recommender system performance by changing the number of training cases, our control task in RTB is more challenging, with much various dynamics from advertising environment such as the fluctuation of market price, auction volume and user behaviour. In {{cite:3d5359ff1dfa2a8080d3a0002fbc2a7297f4fbf1}}, the authors discuss multiple aspects in a performance-driven RTB system, where the impression volume control is one of discussed aspects. Specifically, WL and a model-based controller are implemented to control the impression volume during each time interval. In {{cite:845c24230e66dd8a83371dd49cf375465aca3eb3}}, feedback control is used to perform budget pacing in order to stablise the conversion volume. Compared to {{cite:3d5359ff1dfa2a8080d3a0002fbc2a7297f4fbf1}}, {{cite:845c24230e66dd8a83371dd49cf375465aca3eb3}}, our work is a more comprehensive study focused on the feedback control techniques to address the instability problem in RTB. Besides WL, we intensively investigate the PID controller, which takes more factors into consideration than WL. For the controlled KPIs, we look into the control tasks on both eCPC and AWR, which are crucial KPIs for performance-driven campaigns and branding-based campaigns, respectively. In addition, we proposed an effective model to calculate the optimal eCPC reference to maximise the campaign's clicks using feedback controllers."
] |
[] |
Sentiment Analysis in Scholarly Book Reviews
|
So far different studies have tackled the sentiment analysis in severaldomains such as restaurant and movie reviews. But, this problem has not beenstudied in scholarly book reviews which is different in terms of review styleand size. In this paper, we propose to combine different features in order tobe presented to a supervised classifiers which extract the opinion targetexpressions and detect their polarities in scholarly book reviews. We constructa labeled corpus for training and evaluating our methods in French bookreviews. We also evaluate them on English restaurant reviews in order tomeasure their robustness across the domains and languages. The evaluation showsthat our methods are enough robust for English restaurant reviews and Frenchbook reviews.
|
[
"Aspect-Based Sentiment Analysis consists of several sub tasks. Some studies have proposed different methods for aspect detection and sentiment polarity analysis, others have proposed joint models in order to obtain the aspect and their polarities from the same model, these last models are generally unsupervised.",
"The early work on opinion target detection from on-line reviews presented by {{cite:7c9f3ab9c2ec5b7e09b589b70a1c8ab069f28e7a}} used association rule mining based on Apriori algorithm {{cite:70dfb23426673494e2726888c5ed4889319b25c1}} to extract frequent noun phrases as product features. For polarity detection, they used two seed sets of 30 positive and negative adjectives, then WordNet has been used to find and add the synonyms of the seed words. Infrequent product features or opinion targets had been processed by finding the noun related to an opinionated word.",
"Opinion Digger {{cite:bd69d1ad9ef42b761304d730223ae6b804820469}} also used Apriori algorithm to extract the frequent opinion targets. kNN algorithm is applied to estimate the aspect rating scaling from 1 to 5 stands for (Excellent, Good, Average, Poor, Terrible).",
"Supervised methods use normally Conditional Random Fields (CRF) or Hidden Markov models (HMM). {{cite:d8d4865e37de7aa059654c4f05c518f6cfd11feb}} applied a HMM model to extract opinion targets using the words and their part-of-speech tags in order to learn a model, then unsupervised algorithm for determining the opinion targets polarity using the nearest opinion word to the opinion target and taking into account the polarity reversal words (such as not).",
"A CRF model was used by {{cite:465e61200c10912dbc380190e3e54a73add5d0fe}} with the following features: tokens, POS tags, syntactic dependency (if the opinion target has a relation with the opinionated word), word distance (the distance between the word in the closest noun phrase and the opinionated word), and opinion sentences (each token in the sentence containing an opinionated expression is labeled by this feature), the input of this method is also the opinionated expressions, they use these expressions for predicting the opinion target polarity using the dependency parsing for retrieving the pair target-expression from the training set. We also applied a CRF model with different features {{cite:bc62bc7e64e94c53d998fb9b1fbc536bf64a05ee}}, {{cite:e544354bf447a73043f47864327c6e945b7dd376}}.",
"Unsupervised methods based on LDA (Latent Dirichlet allocation) have been proposed. {{cite:30235ecddc13ffb8072848a614fa29873dc0f524}} used LDA to figure out the opinion targets, determined the number of topics by applying a clustering method, then they used a similar method proposed by {{cite:a304746462a87f475883a2feb4cdf270e45f441f}} to extract the conjunctive adjectives, but not the disjunctive due to the specificity of the domain.",
"{{cite:7042f911bf4ae6c06fdb4bc131d559a40580c899}} proposed Joint model of Sentiment and Topic (JST) which extends the state-of-the-art topic model (LDA) by adding a sentiment layer, this model is fully unsupervised and it can detect sentiment and topic simultaneously.",
"{{cite:0b8ed6cd4d007d30cd039078b4050a0e55b3b074}} modeled the hierarchical relation between product aspects. They defined Sentiment Ontology Tree (SOT) to formulate the knowledge of hierarchical relationships among product attributes and tackled the problem of sentiment analysis as a hierarchical classification problem. Unsupervised hierarchical aspect Sentiment model (HASM) was proposed by {{cite:10eb6f945e171ad24b11d66eec05d1955910b39d}} to discover a hierarchical structure of aspect-based sentiments from unlabeled online reviews."
] |
[] |
A matter of words: NLP for quality evaluation of Wikipedia medical
articles
|
Automatic quality evaluation of Web information is a task with many fields ofapplications and of great relevance, especially in critical domains like themedical one. We move from the intuition that the quality of content of medicalWeb documents is affected by features related with the specific domain. First,the usage of a specific vocabulary (Domain Informativeness); then, the adoptionof specific codes (like those used in the infoboxes of Wikipedia articles) andthe type of document (e.g., historical and technical ones). In this paper, wepropose to leverage specific domain features to improve the results of theevaluation of Wikipedia medical articles. In particular, we evaluate thearticles adopting an "actionable" model, whose features are related to thecontent of the articles, so that the model can also directly suggest strategiesfor improving a given article quality. We rely on Natural Language Processing(NLP) and dictionaries-based techniques in order to extract the bio-medicalconcepts in a text. We prove the effectiveness of our approach by classifyingthe medical articles of the Wikipedia Medicine Portal, which have beenpreviously manually labeled by the Wiki Project team. The results of ourexperiments confirm that, by considering domain-oriented features, it ispossible to obtain sensible improvements with respect to existing solutions,mainly for those articles that other approaches have less correctly classified.Other than being interesting by their own, the results call for furtherresearch in the area of domain specific features suitable for Web data qualityassessment.
|
[
"Automatic quality evaluation of Wikipedia articles has been addressed inprevious works with both unsupervised and supervised learning approaches.The common idea of most of the existing work is to identify a feature set, havingas a starting point the Wikipedia project guidelines, to be exploited with the objective in mind to distinguish Featured Articles.In {{cite:a7d9928908cc9670ad60ca9651decfcb5ee895b9}}, Stvilia et al. identify a relevantset of features, including lingual, structural, historical and reputational aspects of each article. They show theeffectiveness of their metrics by applying both clustering andclassification.As a result, more than 90% of FA arecorrectly identified.",
"Blumenstock {{cite:fc791d0122ef49350ed39c36909bb4f9178acc83}} inspects the relevance of the word-count feature at each quality stage, showing that it can play a very important role in the quality assessment of Wikipedia articles. Only using this feature, the author achieves a F-measure of 0.902 in the task of classifying featured articles and 0.983 in the task of classifying non featured articles. The best results of the investigation are achieved with a classifier based on a neural network implemented with a multi-layer perceptron.",
"In {{cite:1f68aa8cf7f30512cb0927fb39aa19a4457f75f2}}, the authors try to analyze the factors affecting the quality of Wikipedia articles, with respect to their quality class. The authorsevaluate a set of 28 features, over a random sample of 500 Wikipediaarticles, by weighing each metric in different stages using neural networks. Findings are that linguistic features weigh more in the lowest quality classes, and structural features, along with historical ones, become more important as the article quality improves.Their results indicate that the information quality is mainly affected by completeness, and to be “well-written\" is a basic requirement in the initial stage. Instead, reputation of authors or editors is not so important in Wikipedia because of its horizontal structure.In {{cite:278c152aefc822e35b3776c83d6e9832e1ebbcbd}}, the authors consider the quality of the data in the infoboxes of Wikipedia, finding a correlation between the quality of information in the infobox and the article itself.",
"In {{cite:e3e1608ac9dd57535deb13146ee949218caad0d5}}, the authors deal with the problem of discriminatingbetween two large classes, namely NeedWork, GoodEnough (including in GoodEnough both GA and FA), in order to identifywhich articles need further revisions for being featured. They alsointroduce new composite features, those that we have referred to as an “actionable model” in Section .They obtain good classification results, with a F-measure of 0.876 in their best configuration. Theyalso try classification for all the seven quality classes, as done in this work, using arandom forest classifier with 100 trees, with a reduced set offeatures. The poor results (an average F-measure of 0.425) highlightsthe hardness of this fine-grained classification.In this paper, we address this last task in a novel way, by introducing domain features, specially dealing with the medical domain. The results of the investigation are promising.",
"Recent studies specifically address the quality of medical information (in Wikipedia as well as in other resources): in {{cite:7b0c71c0e8b3021949b37bcd8256455ee7de6c98}} and {{cite:ade67b1274eba4a23a501b02f14de495767ea0d4}}, the authors debate if Wikipedia is a reliable learning resource for medical students, evaluating articles on respiratory topics and cardiovascular diseases. The evaluation is carried out by exploiting DISCERNhttp://www.discern.org.uk/, a tool evaluating readability of articles.In {{cite:cd87feb09c5acbfabe7f851da88e19ac86ffa7d0}} the authors provide novel solutions for measure the quality of medical information in Wikipedia, by adopting an unsupervised approach based on the Analytic Hierachy Process, a multi-criteria decision making technique {{cite:2acdc3f6da80632f982a1c504d8d4e15bb891e3d}}.The work in {{cite:bade9f246460a9a9772021f1ce74add47f8db712}} aims to provide the web surfers a numerical indication of Quality of Medical Web Sites.In particular in {{cite:bade9f246460a9a9772021f1ce74add47f8db712}} the author proposes an index to make IQ judgment of the content and of its reliability, to give the so called “surface markers” and “trust indicator”. A similar measurement is considered in {{cite:143ffdec7c0e1eee980531db184844fef56309be}}. where the authors present an empirical analysis that suggests the need to define genre-specific templates for quality evaluation and to develop models for an automatic genre-based classification of health information Web pages. In addition, the study shows that consumers may lack the motivation or literacy skills to evaluate the information quality of health Web pages. Clearly, this further highlights the cruciality to develop accessible automatic information quality evaluation tools and ontologies. Our work moves towards the goal, by specifically considering domain-relevant features and featuring an automatic classification task spanning over more than two classes."
] |
[] |
Bootstrap percolation on geometric inhomogeneous random graphs
|
Geometric inhomogeneous random graphs (GIRGs) are a model for scale-freenetworks with underlying geometry. We study bootstrap percolation on thesegraphs, which is a process modelling the spread of an infection of verticesstarting within a (small) local region. We show that the process exhibits aphase transition in terms of the initial infection rate in this region. Wedetermine the speed of the process in the supercritical case, up to lower orderterms, and show that its evolution is fundamentally influenced by theunderlying geometry. For vertices with given position and expected degree, wedetermine the infection time up to lower order terms. Finally, we show how thisknowledge can be used to contain the infection locally by removing relativelyfew edges from the graph. This is the first time that the role of geometry onbootstrap percolation is analysed mathematically for geometric scale-freenetworks.
|
[
"The GIRG model was introduced in {{cite:d2300f0f1caeddf6d7a2449b6176f1e3289a9681}}, and we rely on many results from this paper. The average distance of a GIRG (which, as we show, agrees with the time until the bootstrap percolation process has infected a constant portion of all vertices) was determined in {{cite:58512968eaafb4cdab1afcc106d0f290c2669c4c}} in a much more general setup.",
"Bootstrap percolation has been intensively studied theoretically and experimentally on a multitude of networks, including trees {{cite:2101015c0020bfc07b7800a947e5aa87f64268d0}}, {{cite:0e0ab6f4b5e5b15227669999d1fd7258b3f4de0f}}, lattices {{cite:2b053fc769350e5799ec0a3043a0387b4cba33ab}}, {{cite:56f8dc1d1c86d154c586da7d7df0cceb66371edd}}, Erdős-Rényi graphs {{cite:baaf2ecef973c44b5ab1ade91dbf90e6dec03564}}, various geometric graphs {{cite:917e3966d7fa8cc5f536f220868518e4e84166a9}}, {{cite:30f8d2c6a6b1ba7abb33133774a6dd576971332b}}, {{cite:06d17c286208137f853d8c31d1098f7a68d3ca63}}, {{cite:ef688a8cbb7289b3cc45b66dbf10828cbdaf6e53}}, and scale-free networks {{cite:a413be7a6ebc2144bd06de39f5f1be1fe8bbe0f4}}, {{cite:afe8f366cd2e3fd181d7ca671593a81af570eb6a}}, {{cite:7d51628575826f2ee5b85ab2eca823a99d9892cc}}, {{cite:335031a6a5503947c1595b67a74652e006bc785b}}. On geometric scale-free networks there are some experimental results {{cite:733963b818394d4ed12d42e8e6d387bc8c131abe}}, but little is known theoretically. Recently, Candellero and Fountoulakis {{cite:0559256898cb91250e912fb60c9906c757a9d303}} determined the threshold for bootstrap percolation on HypRGs (in the threshold case {{formula:2285ca8d-c55d-4e4f-8f4e-21178a49c38c}} , cf. below), but they assumed that the bootstrap is uniformly sampled from all vertices, and not from a geometric region as in our paper. In particular, in this case the bootstrap needs to be polynomial in {{formula:b53122e4-161a-4c20-a1ed-cb1fc5832951}} to start an infection, while in our setting every ball which contains an expected number of {{formula:f74bffb0-6e88-4da7-a735-613d6e136264}} vertices can initiate a large infection whp. Also, due to their setup the authors did not investigate how activity spreads along the geometry of the ground space.",
"While there is plenty of experimental literature and also some mean-field heuristics on other activity spreading processes on geometric scale-free networks (e.g., {{cite:7b6a98e3d1dc13b1f8f452e6740ecceb38352f43}}, {{cite:599d47daf5211efdde391a7857653ad66466bf3a}}, {{cite:db1471ee3d62e6528c6fd62161ac2ca29c4654b3}}, {{cite:7339603ed69755b45ab08d391ff772fe126c46d1}}, {{cite:518861f50dbac11e76890bd4412b12c40911d68b}}, {{cite:8dbec2612cf3b8d747c5bdce246c6770e48bcbb4}}), rigorous mathematical treatments are non-existent with the notable exception of {{cite:5a7761aedec23f48c1211697483def8f06e21d63}}, where rumour spreading is analysed in an SPA model with a push and a push&pull protocol."
] |
[] |
Fair Scheduling for Data Collection in Mobile Sensor Networks with
Energy Harvesting
|
We consider the problem of data collection from a continental-scale networkof energy harvesting sensors, applied to tracking mobile assets in ruralenvironments. Our application constraints favour a highly asymmetric solution,with heavily duty-cycled sensor nodes communicating with powered base stations.We study a novel scheduling optimisation problem for energy harvesting mobilesensor network, that maximises the amount of collected data under theconstraints of radio link quality and energy harvesting efficiency, whileensuring a fair data reception. We show that the problem is NP-complete andpropose a heuristic algorithm to approximate the optimal scheduling solution inpolynomial time. Moreover, our algorithm is flexible in handling progressiveenergy harvesting events, such as with solar panels, or opportunistic andbursty events, such as with Wireless Power Transfer. We use empirical linkquality data, solar energy, and WPT efficiency to evaluate the proposedalgorithm in extensive simulations and compare its performance tostate-of-the-art. We show that our algorithm achieves high data receptionrates, under different fairness and node lifetime constraints.
|
[
"In this section, we review the literature on link scheduling and optimisation in wireless networks. To solve different optimisation goals, recent work considers throughput, energy consumption or time delay.",
"Extensive studies have been conducted on link scheduling in cellular networks. In {{cite:76d570202ce7299ff6154c827b45b4f32888a480}}, the link quality is predicted by an application framework which tracks the direction of travel of mobile phones at the BS. They develop energy-aware scheduling algorithms for different application workloads such as syncing or streaming. Some scheduling optimisations which consider multicast {{cite:1a6c91c933679b112cd6e260c0bf456b85dff896}}, quality-of-service assurance {{cite:fef5663b99133f95500ca2cff8b94f173263a4f1}} and fair relaying with multiple antennas {{cite:ed599ae89492db4ef38d4b43b2b27ae13a1a9c01}} are proposed to achieve optimal delay, capacity gain or network utility.The majority of related work has focused on addressing the scheduling problem in the context of wireless networks {{cite:0994ee510a83a97c3374a723c9bc0112a14a77c3}}, {{cite:e1713b85f601e3036f22daa458f4bef49132e548}}, {{cite:7b0dab8a924e036f74431599e26fcbf8a5100138}}, {{cite:e81f1de593081aed642108bee8fdd7edd186ce32}}. However, the notion of fairness in wireless networks focuses on fair allocation, such as channels, tasks among different queues, or time slots among the links in each super frame, which is different from the fairness in data collection of MSN.",
"A link scheduling for maximum throughput-utility in single-hop networks with the constraint of network delay is presented in {{cite:6961c0c2d373972aaca7de2aab6ec88fabfa9bca}}. It establishes a delay-based policy for utility optimisation. The policy provides deterministic worst-case delay bounds with total throughput-utility guarantee. The author in {{cite:d695d2203c3c2086cd14a60d96667d7e26bb1b36}} proposes an opportunistic scheduling algorithm that guarantees a bounded worst case delay in single-hop wireless networks. However, those scheduling algorithms are not applicable in MSNs, because they do not consider the constraints of energy and fairness of collection. In {{cite:3d33f0b2006eeebe735a4ba66b0fde8c4db71ec8}}, a sensing scheduling among sensor nodes is presented to maximise the overall Quality of Monitoring utility subject to the energy usage. The scheduling algorithm maximises the overall utility which is to evaluate quality of sensor readings based on the greedy algorithm. For body sensor networks, Sidharth, et al. focus on polling-based communication protocols, and address the problem of optimising the polling schedule to achieve minimal energy consumption and latency {{cite:bb7c79355a3527a4e1e5fba3c35446354020fd29}}. They formulate the problem as a geometric program and solve it by convex optimisation.",
"To the best of our knowledge, there is no research focusing on link scheduling optimisation for fair data collection in energy harvesting MSNs. The recent work in the literature is not applicable because they do not optimise the scheduling with the requirements of both energy consumption and data reception fairness. The key difference of our work over previous scheduling optimisation is that for a single-hop MSN which includes a large number of energy harvesting nodes, data collection is maximised in a fair way. We formulate the transmission scheduling optimisation model in Section ."
] |
[] |
Automated Clustering and Program Repair for Introductory Programming
Assignments
|
Providing feedback on programming assignments is a tedious task for theinstructor, and even impossible in large Massive Open Online Courses withthousands of students. Previous research has suggested that program repairtechniques can be used to generate feedback in programming education. In thispaper, we present a novel fully automated program repair algorithm forintroductory programming assignments. The key idea of the technique, whichenables automation and scalability, is to use the existing correct studentsolutions to repair the incorrect attempts. We evaluate the approach in twoexperiments: (I) We evaluate the number, size and quality of the generatedrepairs on 4,293 incorrect student attempts from an existing MOOC. We find thatour approach can repair 97% of student attempts, while 81% of those are smallrepairs of good quality. (II) We conduct a preliminary user study onperformance and repair usefulness in an interactive teaching setting. We obtainpromising initial results (the average usefulness grade 3.4 on a scale from 1to 5), and conclude that our approach can be used in an interactive setting.
|
[
"Automated Feedback Generation.Ihantola et al. {{cite:5e3240fdfa304a2a014aa02a941aa3d875acd370}} present a recent survey oftools for automatic assessment of programming exercises.Pex4Fun {{cite:21d847232e40d2f7a72635a3fe3e684b6a7b8c3d}}, and its successor CodeHunt {{cite:1a0f20567138c79970e8df05eac3c96b4bcc015d}}are browser-based, interactive platformswhere students solve programming assignments with hidden specifications, andare presented with a list of automatically generated test cases.LAURA {{cite:2227878f7798d8c4e0de7bf416911e32ebfb9ac0}} heuristically applies program transformations to student'sprogram and compares it to a reference solution, while reporting mismatchesas potential errors (they could also be correct variations).",
"Trace analysis.Striewe and Goedicke have proposed to create full traces of program behavior automaticallywhile running test cases to make the program behavior visible to students,but the interpretation of the traces is left to the students {{cite:1cfbfc068ede504ef271d2d31626fd4f234b3dba}}.They have also suggested automatically comparing the student's trace to that of a samplesolution {{cite:1d28b74d9d8e0b4ebbe4975271b84f2b7feb0b47}}.However, the approach misses a discussion of the situation when the student's code enters an infinite loop,or has an error early in the program that influences the rest of the trace. In our previous work {{cite:b8da96950d108dda0a7e05b9a59663bdeb0b7c45}} we used a dynamic analysis based approach to finda strategy used by the student, and to generate feedback for performance aspects.However, the approach required specifications manually provided by the teacher,written in a specially designed specification language, and itonly matched specifications to correct attempts, i.e.,it did not involve repair of incorrect attempts.",
"Program Classification in Education.CodeWebs {{cite:a26586ead6b8606ad552b2f92e2d02b684b3ebc1}} classifies different AST sub-trees in equivalence classesbased on probabilistic reasoning and program execution on a set of inputs.The classification is used to build a search engine over ASTs to enablethe instructor to search for similar attempts, and to provide feedback on some class of ASTs.OverCode {{cite:01bf92b4ff724ac27b8ee7364dd3b13e371dd683}} is a visualization system that usesa lightweight static and dynamic analysis, together with manual (provided by the instructor)rewrite rules to group student attempts.Drummond et al. {{cite:9b0fa4b9c92d3593c424b54e5271d394a10c49da}} propose a statistical approach to classifyinteractive programs in two categories (good and bad).",
"Program Repair.The research area of program repair is huge, we mention some of the approaches.Gopinath et al. {{cite:2fe030da74b25aeb4594b05713379a73247aa45f}} propose a SAT-based approach for generating likely bug fixesin programs that manipulate structurally complex data.Könighofer et al. {{cite:605377bf1f93a7c9c04d2688225910c464a959b9}} propose an approach for automatederror localization and correction based on symbolic execution and model-based diagnosis forerror localization, and template-based corrections of the RHS expressions.Another approach {{cite:268546e093b7b8665116c1eaa17062e23c4fc438}}, {{cite:dd4dbfc0b8826becc4905a104353d3eac0610a25}} models the localization and correction problemas a game between the environment that provides different inputs, and system that provides repairs.In contrast, our approach uses dynamic analysis for scalability.These approaches aim to repair large programs, and therefore are not able to generate complex repairs.Our approach repairs small programs in education and uses multiple specifications to find the bestrepair suggestions, and therefore is able to suggest more complex repairsthan correcting RHS expressions.",
"Prophet {{cite:9ae9394ef976fd0c4cd311d9cc736f492520ee7a}} mines database of successful patches and usesthem to repair defects in large, real-world applications.However, it is unclear how this approach would be applicable to educational setting.SearchRepair {{cite:b71cc4a1bbb5e83d562421487bcc1172a82d501d}} mines database of correct code anduses them to repair programs drawn from introductory programming course.However, SearchRepair repairs 49% of programs and the paper does not reporteither the time required for repairs, neither how useful these repairs are in education,so applicability to interactive teaching setting, which is the goal of our work, is questionable.",
"A different approaches are based on program mutation {{cite:5c2fd5efb95057f098903debc24dffee527e0087}},or genetic programming {{cite:a1a3c70b3537b8f63d5a9701965e8323e7e58ba4}}, {{cite:44d80896cb215716a10272b5be90378aae6309c9}} that combinemutation with crossing operators, and choose repairs based on a fitness function.However the search space for mutants is huge;our approach is more systematic in search for repair.",
"Qlose {{cite:9ab7823aaebd57d3e5b1d082687f65775d71fb90}} is approach to automatically repair programs in educationbased on different program distances.The idea to consider different semantic distances is very interesting,however the paper reports only a very small initial evaluation (on 11 programs),and Qlose is able to generate small repairs of RHS expressions, based on simple templates.",
"Program Equivalence.Translation Validation {{cite:7392ee9ee08a4952ecbf728e5fe0197a0a6a3251}}, {{cite:f6f686b6db34b81c52fd6b5bc097b859c20061fa}} is an approach for checkingequivalence of the original program and a result of compilation or optimization,using a simulation-relation between variables of the two programs todemonstrate equivalence, and heuristics and knowledge about performed transformationsto derive a simulation relation.Lahiri et.al. {{cite:a064940f5b7a49782146dad510b38f87a4da8702}} apply automated differential program verificationto show that an approximate program does not diverge significantlyfrom a reference implementation."
] |
[] |
Multimedia Channel Allocation in Cognitive Radio Networks using FDM-FDMA
and OFDM-FDMA
|
In conventional wireless systems, unless a contiguous frequency band withwidth at least equal to the required bandwidth is obtained, multimediacommunication can not be effected with the desired Quality of Service. Wepropose here a novel channel allocation technique to overcome this limitationin a Cognitive Radio Network which is based on utilizing several non-contiguouschannels, each of width smaller than the required bandwidth, but whose sumequals at least the required bandwidth. We present algorithms for channelsensing, channel reservation and channel deallocation along with transmissionand reception protocols with two different implementations based on $FDM-FDMA$and $OFDM-FDMA$ techniques. Simulation results for both these implementationsshow that the proposed technique outperforms the existing first-fit andbest-fit~\cite{b109, b110} allocation techniques in terms of the average numberof attempts needed for acquiring the necessary number of channels for alltraffic situations ranging from light to extremely heavy traffic. Further, theproposed technique can allocate the required numbers of channels in less thanone second with $FDM-FDMA$ ($4.5$ second with $OFDM-FDMA$) even for $96\%$traffic load, while the first-fit and best-fit techniques fail to allocate anychannel in such situations.
|
[
"Multimedia communication through {{formula:bbdf21d6-c832-45fd-a90c-78144955d5ad}} s has already been studied by several authors {{cite:1fc1ef3aa04b3794a0e6af38ddf7947bfa7ee79f}}, {{cite:0a24c4698204e8e32e1b6c7b59ccba50cd50a558}}, {{cite:7699af081414cf0f26b467b26c0c673dd464ef5d}}. Mitola J. first introduced the concept of flexible mobile multimedia communications {{cite:1fc1ef3aa04b3794a0e6af38ddf7947bfa7ee79f}} in a {{formula:9762120b-c228-4689-9a76-0ca2af9f9831}} . Kushwaha et al. {{cite:0a24c4698204e8e32e1b6c7b59ccba50cd50a558}} used fountain coding for packet generation and conversion to send data with high reliability and tolerable delay. Shing et al. {{cite:7699af081414cf0f26b467b26c0c673dd464ef5d}} proposed the idea of dynamic channel selection for video streaming over a {{formula:ae401e0f-ec86-4edc-8e4d-49073fc897fc}} , based on some priority-based scheduling of video signals. On the other hand Lei et. al. worked on spectrum fragmentation by their method “Jello” {{cite:1a09b62711549d85090bd7c1c75703f595835a2c}}, where they detect “edges” of power spectrum, then use classical best-fit, worst-fit and first-fit algorithm for spectrum selection and finally they do a distributed coordinate procedure to synchronize transceiver system. However, in all of these communication schemes, a video signal can not be communicated over the {{formula:f3166a1a-5a0b-478e-a771-a9c452789cad}} unless a channel of sufficiently large bandwidth for maintaining the {{formula:4d9b29b5-53ca-4d43-9851-21837cca5f5c}} of these video signals, is allocated from the white spaces of the spectrum. Thus, even if the sum of all the available white spaces in the spectrum may be larger than the required bandwidth for transmitting a video signal, it may not be possible to transmit the video signal if there is no single white space in the spectrum which can provide the required large bandwidth for its communication. Basically, this is a situation of fragmentation of the spectrum into small holes, with no hole being large enough to accommodate a video signal transmission. It was mentioned by Akyildiz et al. {{cite:36c4b81f4b259ff3c8f0ad4ccff48216e149e675}} that “CR users may not detect any single spectrum band to meet the user's requirements. Therefore, multiple noncontiguous spectrum bands can be simultaneously used for transmission in CR networks”. Some authors have addressed this implementation issue of the proposal by using Orthogonal Frequency Division Multiplexing ({{formula:5618f5ed-3791-48e6-acbe-f82a8da93b3e}} ) - based {{formula:7e4fd73c-0e96-47f0-870e-c7a3a18549c1}} {{cite:7f3fb52a340a3c8dc3e41f971667bdbcce45da59}}, {{cite:12317ef0dc98fc5c2c1f1adbfc16206ba75b91a1}}. However, Multi-Band {{formula:defb8795-3d06-48cf-a72b-9157435b9657}} ({{formula:01cfd457-7f79-4f3a-8a65-7a67144f4f46}} ) system for allowing more than one sender to send their messages in the {{formula:de25e51b-85bd-4f13-8587-d7df03a6d55c}} is still a challenging problem {{cite:7acb1d3f861277b2db7a5abd1c431881e58b4322}}.",
"Techniques for detection of unused spectrum and sharing the spectrum without harmful interference with other users with the help of a Common Control Channel ({{formula:aa347959-311c-4457-8dd2-ed67287824ed}} ) have been presented by Krishnamurthy et al. {{cite:4c922415ccccc0e67b5cd9f4081b71d02c05c972}}, Masri et al. {{cite:c8a10ec67c9309b51d90c052b01656877e37ea16}} and Bayhan and Alagöz {{cite:dd349d9df36fc48fcef1c41cde60030e0dc2c36c}}. The {{formula:c167f1b1-1650-4f81-b7ca-07901af96751}} is used for supporting the transmission coordination and spectrum related information exchange between the {{formula:2f8f5ca7-e0b9-4e19-b4d7-06ff0546e9be}} users. It facilitates neighbor discovery, helps in spectrum sensing coordination, control signaling and exchange of local measurements between the {{formula:7709947a-de80-404b-9928-d67df5b9cf17}} users. Spectrum sensing without using a {{formula:f64d4fec-f4e8-40b2-b4a8-40ad43b16cb4}} has been considered by Kondareddy et al. {{cite:41b5bcd662bac0941401655cbb2f8b603fae60a8}} and Xin and Cao {{cite:b8a3e09af45d8ed4d0fa50a25ec8dcb38c873f8f}}.",
"Taxonomy, open issues, and challenges for channel assignment algorithms in a {{formula:9b803f79-1c1f-4abb-95a6-bc65437318e5}} have been described in {{cite:4331e20f6fffb0d129c4619290e0b14702f73869}}. Allocation schemes can be fixed {{cite:dd4841bd26a68fa8294bb38a9fecb17c1b1b9704}}, {{cite:60e061a89f5369aeac31536530d42867a8a174bb}}, dynamic {{cite:3e3bc93afbe0f531208680457d60263452013288}}, {{cite:e47767f9963da36d8c8a4547c191c6684e8e36d7}}, {{cite:7699af081414cf0f26b467b26c0c673dd464ef5d}}, {{cite:dd4841bd26a68fa8294bb38a9fecb17c1b1b9704}}, {{cite:60e061a89f5369aeac31536530d42867a8a174bb}} or {{cite:dd4841bd26a68fa8294bb38a9fecb17c1b1b9704}}, {{cite:60e061a89f5369aeac31536530d42867a8a174bb}} depending on the flexibility of assigning channels to the cells in the network. The dynamic channel allocation in the spectrum is similar to the computer classical memory management strategies like “first-fit”, “best-fit”, and “worst-fit” {{cite:fb007dbc710810df736cbad9d38b7318f111cdac}}. Very recently, Bayhan and Alagöz {{cite:dd349d9df36fc48fcef1c41cde60030e0dc2c36c}}, {{cite:2e16dee13c3700aa1c5286777b48398dc993dbf8}} have proposed best-fit channel selection techniques in cognitive radio networks."
] |
[] |
Feature Selection as a Multiagent Coordination Problem
|
Datasets with hundreds to tens of thousands features is the new norm. Featureselection constitutes a central problem in machine learning, where the aim isto derive a representative set of features from which to construct aclassification (or prediction) model for a specific task. Our experimentalstudy involves microarray gene expression datasets, these are high-dimensionaland noisy datasets that contain genetic data typically used for distinguishingbetween benign or malicious tissues or classifying different types of cancer.In this paper, we formulate feature selection as a multiagent coordinationproblem and propose a novel feature selection method using multiagentreinforcement learning. The central idea of the proposed approach is to"assign" a reinforcement learning agent to each feature where each agent learnsto control a single feature, we refer to this approach as MARL. Applying thisto microarray datasets creates an enormous multiagent coordination problembetween thousands of learning agents. To address the scalability challenge weapply a form of reward shaping called CLEAN rewards. We compare in total ninefeature selection methods, including state-of-the-art methods, and show thatthe proposed method using CLEAN rewards can significantly scale-up, thusoutperforming the rest of learning-based methods. We further show that a hybridvariant of MARL achieves the best overall performance.
|
[
"Traditionally, there are three categories of feature selection algorithms. Filters {{cite:82fe30eb04e22a21d02a705f7c8d6fcd02608104}} use variable ranking techniques that rely on statistical measures (such as Pearson's correlation coefficient and mutual information) to select top-ranked features. It is important to note that filtering is performed as a data pre-processing step i.e. independent of the choice of the classification algorithm. Filters have been demonstrated to perform well in many cases {{cite:d64bdf470642bbf5307fdd8dc5a90ec2e4e4a92e}} and they are also computationally cost-effective and fast {{cite:d64bdf470642bbf5307fdd8dc5a90ec2e4e4a92e}}.",
"Let {{formula:ccf57128-f84d-4fb5-9ec0-eb6c30360f1d}} and {{formula:9659a2eb-7205-46f1-837b-372c4eefcee2}} be the training set with the class labels vector. When each feature {{formula:6bcbb054-5d56-45af-9c73-80120a73fe49}} is considered individually we refer to it as a univariate filter. An example of this is the univariate correlation-based feature selection (uCFS) {{cite:82fe30eb04e22a21d02a705f7c8d6fcd02608104}} where the Pearson's correlation between each feature space {{formula:327c9b84-2367-40c2-b4b9-e536937000fd}} and the target class {{formula:7015633f-2a5e-4dab-aa3d-2fc08f8f8c12}} is taken. Based on the square of this score the top-ranked features are selected; the square is considered so negatively correlated features are included.",
"When a subset of features is considered we refer to it as a multivariate filter. A popular example of this is the minimal-redundancy maximal-relevance (mRMR) method {{cite:c361424baccf7b0385f7538208528dc6428ecb7c}}. mRMR searches for a subset {{formula:c6bb0541-2d56-4f08-864e-a69d9d73e6d9}} that maximises relevance by maximising the mutual information between each individual feature space {{formula:f5c66139-6448-4557-8dca-01840b1b1647}} and the class {{formula:4609d633-ee80-4993-894c-04c4776feb92}} , and at the same time minimises redundancy by minimising the mutual information between any two feature spaces {{formula:31f7f32a-6347-42c5-8521-e08c226d7d6f}} and {{formula:ba9711fb-01b7-416b-8642-8026fe5047ef}} {{formula:fa9a64cd-8261-4790-986f-b93440e87c0a}} within the subset. Another example of this category is the multivariate CFS {{cite:2f7a7b53f60015de5ee3eb79ecf4df4e661f6224}}.",
"Wrappers {{cite:d64bdf470642bbf5307fdd8dc5a90ec2e4e4a92e}} consider the classification algorithm's performance (e.g. accuracy) as the objective function in order to evaluate feature subsets. This is known to be NP-hard as enumeration of all {{formula:0a847b39-5fd9-45fd-b739-52706c2afe00}} possible feature subsets is intractable; for this reason heuristic search algorithms are employed. Wrappers typically outperform filters because they take into account the feedback from the classifier, at the expense of being computationally intensive and slow. Examples include hill climbing algorithms {{cite:2f7a7b53f60015de5ee3eb79ecf4df4e661f6224}} such as forward selection and backward elimination, best-first search {{cite:2f7a7b53f60015de5ee3eb79ecf4df4e661f6224}}, sequential floating forward selection {{cite:d64bdf470642bbf5307fdd8dc5a90ec2e4e4a92e}}, Monte Carlo tree search {{cite:276e7e4645e983c3e9e0afbd030fb27d0d9f0bec}}, neural networks {{cite:7024e40db1088fadb878cbc4624c82286afd1e9d}} and genetic algorithms {{cite:af5cd7914745d6f67c5f06fc054d6679f8599c6a}}, {{cite:d64bdf470642bbf5307fdd8dc5a90ec2e4e4a92e}}, {{cite:d52c60c77fb69ddc401f2bf4a8ee22e9d262765f}}.",
"A hybrid filter-wrapper approach, where a filter is first used to bring down the number of features and then a wrapper is applied, combines the advantages of both worlds. Alternatively, in embedded {{cite:82fe30eb04e22a21d02a705f7c8d6fcd02608104}}, {{cite:d64bdf470642bbf5307fdd8dc5a90ec2e4e4a92e}} approaches, the feature selection process is built-in or “embedded” in the classifier directly; popular examples are decision trees {{cite:025819ed141a5d81f4dd9be803d0bb8785155806}} and regularization methods {{cite:773b02838ecd06fc034dbe6245f51fb8a3ae785a}}."
] |
[] |
Bank distress in the news: Describing events through deep learning
|
While many models are purposed for detecting the occurrence of significantevents in financial systems, the task of providing qualitative detail on thedevelopments is not usually as well automated. We present a deep learningapproach for detecting relevant discussion in text and extracting naturallanguage descriptions of events. Supervised by only a small set of eventinformation, comprising entity names and dates, the model is leveraged byunsupervised learning of semantic vector representations on extensive textdata. We demonstrate applicability to the study of financial risk based on news(6.6M articles), particularly bank distress and government interventions (243events), where indices can signal the level of bank-stress-related reporting atthe entity level, or aggregated at national or European level, while beingcoupled with explanations. Thus, we exemplify how text, as timely, widelyavailable and descriptive data, can serve as a useful complementary source ofinformation for financial and systemic risk analytics.
|
[
"The automatic identification of events in chronological text such as news has been explored at least since the 90s, when a DARPA-coordinated effort was organized {{cite:f828d96255086f8a240755221aee1eeb85fc0fa2}} that set the foundation for what is known as topic detection and tracking (TDT), where news streams are analyzed in order to identify reporting on new events as well as recurring reporting relating to earlier events. The early detection and tracking methods were data-driven, based on clustering in particular, and intended to capture any kind of event (see, e.g., {{cite:cdd3965934016b62da0763ea39b0f5c8539350b3}}).",
"A related area of research that since has emerged, mainly stemming from the area of information extraction, is event extraction, which aims at extracting complex structured information about events in terms of pre-defined types of events and entities, as well as attributes of events and roles of entities (cf., e.g., {{cite:bc49935b35c96ab0500b777bb206cd16719f4474}}).The event extraction techniques focus on identifying and extracting more specific types of information, with explicit semantic interpretation, in contrast to the TDT approach. As the information of interest is often particular to an expert domain and task, the techniques tend to require substantial expert guidance in terms of designing linguistic patterns or annotating text, which makes them less applicable in new domains where fewer resources may be available to target specific tasks and the information of interest may be difficult to strictly define.Efforts focusing on the financial domain and identification of specific types of risk include {{cite:0db2ea3682bdd5e364aad190d054608ebc2dfea5}}, {{cite:1ad51841f9999c17c0833e0873850f78dd0f2ccb}} and {{cite:862b79880d0069ae865e13d564b0b85b35934012}}. Tanev et al. {{cite:4613d808eee25f71b8f0cd7a16c5c6f01f2254d2}} also explore the combination of data-driven preprocessing with the knowledge-driven approach to extracting events, as they monitor violent and disaster events in news. Hogenboom et al. {{cite:2aac749c70b2d0dd916bd887699f6db34a461956}} provide a thorough overview of how event extraction has evolved in various fields.",
"Parallel to this view on event discovery, which naturally places description of events at its heart, non-text data sources have also been investigated for the detection of significant events, or the risk thereof, such as failure of companies using machine learning {{cite:10b3dc5cb31d64acf3b189aa9fa13178fc51f269}}, {{cite:b1cedba8fa3da48ab7a395cdb47aa2a952a7014d}}. The focus is then primarily on estimating the likelihood that a particular type of event will occur. While the specification of events in text mining tends to be more idiosyncratic to the input data, the events in distress prediction tend to be specified by when they occur and what entities they involve, as is the case in this paper, too. Such event specifications are easier to recombine with new data, including text data given appropriate modeling.",
"In particular, prediction of bank distress has been a major topic both before and following the global financial crisis. Many efforts are concerned with identifying thebuild-up of risk at early stages, often relying upon aggregatedaccounting data to measure imbalances (e.g., {{cite:2ebc85ddf6dc98ec5cb736c5a52de89740962909}}, {{cite:8bc42110e8d3fa240abdb1c540fba52e948e4b86}}, {{cite:cfe569c3bbde64e11d5a61aa95168fd6b66d5490}}).Despite their rich information content, accounting data pose majorchallengesdue to restricted access, as well as low reporting frequency and long publication lags.A widely available and more timely source of information is the use of market data to indicateimbalances, stress and volatility (e.g., {{cite:f4b9f9a271ebe28cbe739a9d4f6d66077550dfaa}}, {{cite:d32bfdba6e74c2c4c8fbdbe6baa454ef8f0526ff}}).Yet, market prices provide little or no descriptive information perse, and only yield information about listed companies or companies'traded instruments (such as Credit Default Swaps). This points tothe potential value of text as a source for understanding events such as bank distress.More generally, central banks are starting to recognize the utility of text data in financial risk analytics, too. {{cite:a83f1854c68b99ab44626ccb8cb77599848195e9}}, {{cite:51c73637e2f936aee5afba0ea058d942c955db2f}}",
"The literature on text-based computational methods for measuring riskor distress is still rather scarce and scattered. For instance, Nyman et al. {{cite:24520531ef11c38edd05a9fe22813be881ba9ec5}}analyze sentiment trends in news narratives in terms of excitement/anxietyand find increased consensus to reflect pre-crisis market exuberance,Soo {{cite:1f2eefab0f0f5bf7c4b997d3d46cd9c349007f3f}} analyzes the connection betweensentiment in news and the housing market and Cerchiello et al. {{cite:103ce6c0265bb751b41c5c8cf5ec4c95fc9c6eaa}} analyse bank risk contagion with both market prices and sentiment index. All three approaches rely on manually-crafteddictionaries of sentiment-bearing words. While such analysis can provideinteresting insight as early work on processing expressions intext to study risk, the approach is generally limiting as dictionaries are cumbersometo adapt to specific tasks, incomplete and unable to handle semantics beyond single words well. Nevertheless, sentiment analysis based on such simple approaches works quite well due to the fact that it relies on human emotions as strong priors in a way that generalizes across tasks and data, and because lower recall may be countered by the scale of the data.",
"Malo et al. {{cite:c1a015ca81e12937a25de7db5a4b147060d05445}} explore a linguistically more sophisticated approach that models financial sentiment compositionally, although without semantic generalization, supervised by a custom data set of annotated phrases. Hogenboom et al. {{cite:862b79880d0069ae865e13d564b0b85b35934012}} integrate their linguistically aware event extraction techniques with the conventional Value at Risk model to account for certain cases of event-driven market effects.",
"Data-driven approaches, such as Wang & Hua {{cite:caf16c79b3579442ed24265cbea29d54647d511b}}predicting volatility of company stocks from earning calls, may avoidthe issues of handcrafted features and manually annotated corpora. Their method, although allegedly providing good predictiveperformance gains, offers only limited insight into the risk-relatedlanguage of the underlying text data. It also leaves room for furtherimprovements with regard to the semantic modeling of individual wordsand sequences of words, which we address. Further, Lischinsky {{cite:c8963ff502aa6c924b324bc985f2961649caf7ca}}performs a crisis-related discourse analysis of corporate annual reportsusing standard corpus-linguistic tools, including some data-drivenmethods that enable exploration based on a few seed words. His analysisfocuses extensively on individual words and their qualitative interpretationas part of a crisis discourse, which likewise provides rather limited insight compared to what full sentences are able to communicate.Finally, Rönnqvist & Sarlin {{cite:e278e2f7794f15073c18b4175c787fb19d43121e}}construct network models of bank interrelations based on co-occurrencein news, and assess the information centrality of individual bankswith regard to the surrounding banking system, a fully data-driven approach that could be further enhanced by semantic modeling and conditioning.",
"In the following, we introduce the deep learning approach and our particular model, along with further relevant previous work."
] |
[] |
Consistently faster and smaller compressed bitmaps with Roaring
|
Compressed bitmap indexes are used in databases and search engines. Manybitmap compression techniques have been proposed, almost all relying primarilyon run-length encoding (RLE). However, on unsorted data, we can get superiorperformance with a hybrid compression technique that uses both uncompressedbitmaps and packed arrays inside a two-level tree. An instance of thistechnique, Roaring, has recently been proposed. Due to its good performance, ithas been adopted by several production platforms (e.g., Apache Lucene, ApacheSpark, Apache Kylin and Druid). Yet there are cases where run-length encoded bitmaps are smaller than theoriginal Roaring bitmaps---typically when the data is sorted so that thebitmaps contain long compressible runs. To better handle these cases, we builda new Roaring hybrid that combines uncompressed bitmaps, packed arrays and RLEcompressed segments. The result is a new Roaring format that compresses better. Overall, our new implementation of Roaring can be several times faster (up totwo orders of magnitude) than the implementations of traditional RLE-basedalternatives (WAH, Concise, EWAH) while compressing better. We review thedesign choices and optimizations that make these good results possible.
|
[
"There are many RLE-based compression formats. For example, WAH organizes the data inliteral and fill words.Literal words contain a mix of {{formula:dd592fa5-6058-4f6a-aa0c-df8cf62f67e8}} zeros and ones (e.g., {{formula:a3b9537a-b53d-482d-a362-ebc26c28769a}} )where {{formula:22dceb98-7433-4d61-b7fe-532275239411}} denotes the word size in bits: typically {{formula:755ccece-d750-4a4b-9566-cc20c5c23656}} or {{formula:1459d4f1-cec3-475c-a65e-19ae0bb785dd}} .Fill words are made of just {{formula:2c46167f-542a-4ef4-b633-7f12a38c225c}} ones or just {{formula:4089da40-7677-4298-88aa-c266c20a127e}} zeros (i.e.,{{formula:fe08eb35-8f49-4b03-af5c-e1cb4ea6deba}} or {{formula:8ef52049-08f1-4d1e-b24b-627991b1f11d}} ).WAH compresses sequences of consecutive identical fill words.The most significant bit of each word distinguishes betweenfill and literal words. When it is set to one, the remaining {{formula:cec1c582-f069-484e-9efb-4e4065e2a10a}} bits store the {{formula:781ffe1d-d27b-43d6-9347-b2bcf51d0785}} bits of a literal word. When it is set to zero, the second most significant bit indicates the bit value whereas the remaining bits are used to store the number of consecutive identical fill words (the run length).Concise is a variation that reduces the memory usage when the bitmap ismoderatelysparse {{cite:1118f74bfd4a3bc8ddc38132a3a71752b6445501}}. Instead of storing the run lengthusing {{formula:c40f11d3-8595-4713-ab6e-b45016da7d71}} bits, Concise uses only {{formula:1bb61f94-4867-4712-88fe-528808dd3006}} bits to indicate a run length {{formula:dd6bc07e-6830-41f3-9b86-34516370bee8}} , reserving {{formula:618c129f-7c72-4479-9916-7bdb6e2971c0}} bits to store a value {{formula:855731cd-2cdd-4c2f-82b2-93c1f4b2de98}} .When {{formula:c859820f-1f34-4cc2-9d73-660aa3c73dbc}} is non-zero, we decode {{formula:a785bd8f-5a76-4428-b685-7eff1d32eb40}} fill words, plus a single {{formula:f870b335-6f3f-422d-b287-fcdf12b5b271}} bit word with its {{formula:529acf07-b208-49a6-9c8a-960803fe74cc}} bit flipped.Whereas WAH would use 64 bits per value to store the set {{formula:132397ea-9864-466d-be28-de3627082f99}} , Concise would only use 32 bits per value.",
"EWAH is similar to WAH except thatit uses a marker word that indicates the number of fill words to follow, their type, as well as the number of literal words to follow.Unlike WAH and Concise, which represent the bitmap as a series of {{formula:3cb67d53-e3d6-4fa9-a1d3-46312c621fed}} -bit words, EWAH uses {{formula:0c113e79-94e2-44d9-877d-c60a85c81a3c}} -bit words.The EWAH format {{cite:323bf59620caadef621d94588942b9a42652e11a}} supports a limited form of skipping becauseit uses marker words to record the length of the sequences offill and literal words.For EWAH,if there are long sequences ofliteral words, one does not need to access them allwhenseeking data that is further along. Guzun etal. {{cite:bda004685723d9d7617f077892b43e88450472e2}}, {{cite:4b5e1303a6c61f527e7d1d7ded9380c39d35f430}} found that EWAH offers better speed than WAH and Concise, and our own experiments support this observation (see § ).",
"Beside WAH, Concise and EWAH, there are many other similar alternatives. We refer the interested reader to Chen et al. {{cite:4b3206db069e572de8173e34e7fee178121685d4}}, who review over a dozen compressed bitmap formats.",
"The general idea behind Roaring—using different container types depending on the data characteristics—is not novel. It is similar to O'Neil and O'Neil's RIDBit external-memory system: a B-tree of bitmaps, where a list is used instead when the density of a chunk is too small {{cite:546305d2e2807a36fabf37fec3e708e7bd16ae52}}, {{cite:488ef4628864726065391d7237f1fe35f441e6e8}}. Similarly,Culpepper and Moffat proposed a hybrid inverted index (hyb+m2) where some sets of document identifiers are stored as compressed arrays whereas others are stored as bitmaps {{cite:795c868b5f64f4087f51e7cecee6c31e0cdf2bb6}}.To reduce cache misses, Lemire et al. partitioned the Culpepper-Moffat index into chunks that fit in L3 cache {{cite:b674ce3424c26586f03d7bca38b75512354df83b}}, effectively creating sets of bitmaps and arrays.Roaring is also reminiscent of C-store where different columns are stored in different formats depending on their data characteristics (number of runs, number of distinct values) {{cite:ddc36f46a07f6f06d8bcdc1fb89a964da19142e3}}.It is likely that we could find many other similar instances."
] |
[] |
Semi-supervised Word Sense Disambiguation with Neural Models
|
Determining the intended sense of words in text - word sense disambiguation(WSD) - is a long standing problem in natural language processing. Recently,researchers have shown promising results using word vectors extracted from aneural network language model as features in WSD algorithms. However, a simpleaverage or concatenation of word vectors for each word in a text loses thesequential and syntactic information of the text. In this paper, we study WSDwith a sequence learning neural net, LSTM, to better capture the sequential andsyntactic patterns of the text. To alleviate the lack of training data inall-words WSD, we employ the same LSTM in a semi-supervised label propagationclassifier. We demonstrate state-of-the-art results, especially on verbs.
|
[
"The development of large lexical resources, such as WordNet {{cite:3110512b951f2cc90fccd2a47564b1617623fd39}} and BabelNet {{cite:7bd69a98f39dbe61b1ca7f962622024b8f59beb1}}, has enabled knowledge-based algorithms which show promising results on all-words prediction tasks {{cite:099f974c4f97a15e952aefc66996944a4c75bde5}}, {{cite:9bca2a6b1083d086e44b27f3b13eebb144ad18bd}}, {{cite:cb83cded8a998e29a436e39650acaa0d10bb71a2}}.WSD algorithms based on supervised learning are generally believed to perform better than knowledge-based WSD algorithms, but they need large training sets to perform well {{cite:0710145af35f7d9fa5c2b0c9ef010eadf4a2b6bb}}, {{cite:faeaeb75f65edcd150579f5d8400ac0e78068747}}, {{cite:51bd4bd5795391b86710f56c06a7cb878af0f149}}, {{cite:917051b953d5658e5274063f10402067ad3f127a}}.Acquiring large training sets is costly.In this paper, we show that a supervised WSD algorithm can perform well with {{formula:1db3b8da-19b7-476f-b81a-679dc3829191}} training examples per sense.",
"In the past few years, much progress has been made on using neural networks to learn word embeddings {{cite:d4e581f2b338556a5dad03c7af8a70ace015d898}}, {{cite:e2b3591c15e4cd848df525b04eb445b0cb4266c9}}, to construct language models {{cite:4ffc4e106b5629221193445379b4912bae33d5bf}}, perform sentiment analysis {{cite:c3e8a131804c999808feb465c7349ee3eb382c05}}, machine translation {{cite:f9c986bf5b4777112d352268308c1d1829876565}} and many other NLP applications.",
"A number of different ways have been studied for using word embeddings in WSD. There are some common elements:",
"[noitemsep,nolistsep]Context embeddings. Given a window of text {{formula:16008b7d-640a-434b-8bcd-6b4e9ad0636e}} surrounding a focus word {{formula:40c0c35a-6e1b-44f3-83d6-f5d320ff7cde}} (whose label is either known in the case of example sentences or to be determined in the case of classification), an embedding for the context is computed as a concatenation or weighted sum of the embeddings of the words {{formula:05476c02-f5c7-4668-80a7-24c9863602e4}} . Context embeddings of various kinds are used in both {{cite:3eaf3163c8168bd1e05ac45ff0844b3080eb5830}} and {{cite:ddde758c48a90fc7c8566bd69f74c020d7d6b101}}.Sense embeddings. Embeddings are computed for each word sense in the word sense inventory (e.g. WordNet). In {{cite:a071ff4956a37a930a485479dd4c7c354f4d6518}}, equations are derived relating embeddings for word senses with embeddings for undisambiguated words. The equations are solved to compute the sense embeddings. In {{cite:3eaf3163c8168bd1e05ac45ff0844b3080eb5830}}, sense embeddings are computed first as weighted sums of the embeddings of words in the WordNet gloss for each sense. These are used in an initial bootstrapping WSD phase, and then refined by a neural network which is trained on this bootstrap data.Embeddings as SVM features. Context embeddings {{cite:ddde758c48a90fc7c8566bd69f74c020d7d6b101}}, {{cite:a6871d92a97269f27438b67d656e82202c34e80c}}, or features computed by combining context embeddings with sense embeddings {{cite:a071ff4956a37a930a485479dd4c7c354f4d6518}}, can be used as additional features in a supervised WSD system e.g. the SVM-based IMS {{cite:917051b953d5658e5274063f10402067ad3f127a}}. Indeed Iacobacci et al. supervisedword2vec found that using embeddings as the only features in IMS gave competitive WSD performance.Nearest neighbor classifier. Another way to perform classification is to find the word sense whose sense embedding is closest, as measured by cosine similarity, to the embedding of the classification context. This is used, for example, in the bootstrapping phase of {{cite:3eaf3163c8168bd1e05ac45ff0844b3080eb5830}}.Retraining embeddings. A feedforward neural network can be used to jointly perform WSD and adjust embeddings {{cite:3eaf3163c8168bd1e05ac45ff0844b3080eb5830}}, {{cite:a6871d92a97269f27438b67d656e82202c34e80c}}.",
"In our work, we start with a baseline classifier which uses 1000-dimensional embeddings trained on a 100 billion word news corpus using Word2Vec {{cite:d4e581f2b338556a5dad03c7af8a70ace015d898}}. The vocabulary consists of the most frequent {{formula:b52ca5a8-fdae-451d-a24d-b8231b9eb81c}} words, without lemmatization or case normalization. Sense embeddings are computed by averaging the context embeddings of sentences which have been labeled with that sense. To classify a word in a context, we assign the word sense whose embedding has maximal cosine similarity with the embedding of the context. This classifier has similar performance to the best classifier in {{cite:ddde758c48a90fc7c8566bd69f74c020d7d6b101}} when SemCor is used as a source of labeled sentences. The Word2Vec embeddings are trained using a bag of words model, i.e. without considering word order in the training context, and word order is also not considered in the classification context. In Section we show that using a more expressive language model which takes account of word order yields significant improvements.",
"Semi-supervised learning has previously been applied successfully to word sense disambiguation. In {{cite:beea4fd77495d79cd7bf83176e6c8a2c53f1ee24}} bootstrapping was used to learn a high precision WSD classifier. A low recall classifier was learned from a small set of labeled examples, and the labeled set then extended with those sentences from an unlabeled corpus which the classifier could label with high confidence. The classifier was then retrained, and this iterative training process continued to convergence. Additional heuristics helped to maintain the stability of the bootstrapping process. The method was evaluated on a small data set.",
"In {{cite:cb7525858eea118d61f65c11f9c20098c3f7dbaf}}, a label propagation algorithm was proposed for word sense disambiguation and compared to bootstrapping and a SVM supervised classifier. Label propagation can achieve better performance because it assigns labels to optimize a global objective, whereas bootstrapping propagates labels based on local similarity of examples.",
"In Section we describe our use of label propagation to improve on nearest neighbor for classification."
] |
[] |
Quantum Recommendation Systems
|
A recommendation system uses the past purchases or ratings of $n$ products bya group of $m$ users, in order to provide personalized recommendations toindividual users. The information is modeled as an $m \times n$ preferencematrix which is assumed to have a good rank-$k$ approximation, for a smallconstant $k$. In this work, we present a quantum algorithm for recommendation systems thathas running time $O(\text{poly}(k)\text{polylog}(mn))$. All known classicalalgorithms for recommendation systems that work through reconstructing anapproximation of the preference matrix run in time polynomial in the matrixdimension. Our algorithm provides good recommendations by sampling efficientlyfrom an approximation of the preference matrix, without reconstructing theentire matrix. For this, we design an efficient quantum procedure to project agiven vector onto the row space of a given matrix. This is the first algorithmfor recommendation systems that runs in time polylogarithmic in the dimensionsof the matrix and provides an example of a quantum machine learning algorithmfor a real world application.
|
[
"The development of quantum algorithms for linear algebra was initiated by the breakthrough algorithm of Harrow, Hassidim, Lloyd {{cite:70dbaf1df721bdbf490ac8151ccf56385a698e0c}}. The HHL algorithm takes as input a sparse (the number of non zero entries in each row of the matrix is polylogarithmic) and well-conditioned system of linear equations and in time polylogarithmic in the dimension of the system outputs a quantum state which corresponds to the classical solution of the system. Note that this algorithm does not explicitly output the classical solution, nevertheless, the quantum state enables one to sample from the solution vector. This is a very powerful algorithm andhas been very influential in recent times, where several works {{cite:4bb68bd2967aeb35bae72affd64ea5fa6a90ade9}}, {{cite:afdc80d8e26692deb56fd76f616c2ce4f3196789}}, {{cite:957a32191852c7eac9c6c039555620d0a65006b8}} obtained quantum algorithms for machine learning problems based on similar assumptions. However, when looking at these applications, one needs to be extremely careful about two things: first, the assumptions that one needs to make on the input in order to achieve efficient running time, since, for example, the running time of the HHL algorithm is polylogarithmic only when the matrix is well conditioned (i.e. the minimum singular value is at least inverse polynomially big) and sparse; and second, whether the quantum algorithm solves the original classical problem or a weaker variant to account for the fact that the classical solution is not given explicitly but is encoded in a quantum state {{cite:a7f37575bda7c3682d0c9dfbb1f7c5f1ad955594}}, {{cite:957a32191852c7eac9c6c039555620d0a65006b8}}. In addition, we mention a recent but orthogonal proposal to use techniques inspired by the structure of quantum theory for classical recommender systems {{cite:5e33b5578010488e57698e2436fb150117f9eecd}}.",
"Let us be more explicit about our algorithm's assumptions. We assume the data is stored in a classical data structure which enables the quantum algorithm to efficiently create superpositions of rows of the subsample matrix. The HHL algorithm also needs to be able to efficiently construct quantum states from classical vectors given as inputs. In the Appendix,we describe a classical data structure for storing the matrix {{formula:5209a318-2341-49de-9e0e-d298ad03cb3c}} . The data structure maintains some extra information about the matrix entries, so that, the total memory needed is linear (up to polylogarithmic terms) in the number of entries in the subsample matrix, the data entry time remains polylogarithmic in the matrix dimensions, and an algorithm with quantum access to the data structure can create the necessary superpositions in polylogarithmic time. Note also, that even in the case the data has been stored as a normal array or list, we can preprocess it in linear time to construct our needed data structure. Thus, our quantum algorithm works under the same memory model as any other quantum query algorithm (e.g. Grover's algorithm): it assumes that there exists a classical data structure to which we can make quantum queries.Overall, our system retains the necessary properties for the data entry and retrieval stage. Moreover, the classical complexity of matrix reconstruction does not change given the new data structure.",
"Importantly, in our system, we do not perform any preprocessing nor do we need any extra memory. Our recommendation algorithm just performs an online computation that requires time {{formula:abea9c1d-406b-4728-a879-1933060cee34}} . This can be viewed as exponentially smaller than the classical time if the rank {{formula:ccf37ec6-eb50-464b-95a6-df14835481d9}} is a smallconstant and the matrix dimensions are of the same order.Unlike the HHL algorithm, our running time does not depend on the sparsity of the input matrix nor on its condition number, i.e. its smallest singular value. In other words, we do not make any additional assumptions about the classical databeyond the low rank approximation assumptions made by classical recommendation systems.",
"It is also crucial to note that we have not changed what one needs to output, as was the case for the HHL algorithm and its applications, where instead of explicitly outputting a classical solution, they construct a quantum state that corresponds to this solution. We have instead described a real world application, where the ability to sample from the solution is precisely what is needed.",
"The rest of this paper is organized as follows. We introduce some preliminaries in section 2. In sections 3 and 4we show that sampling from an approximate reconstruction of the matrix {{formula:cba45ec5-3ff9-44c1-b144-6898e05b8742}} suffices to provide good recommendations and that if the sub-samples {{formula:8defcc2b-7baa-4e6f-acb6-a080daf93190}} are uniformly distributed then projecting ontothe top {{formula:d620ebc5-b543-4cde-a5f2-772ccc4cf1af}} singular vectors of {{formula:5f61d6d4-8ec8-4cb3-a731-5c4c9b8e6e8d}} is an approximate reconstruction for {{formula:8773859a-db6f-45c7-827a-c7f00e719310}} . In section 5 we describean efficient quantum algorithm for projecting a vector onto the space of singular vectors of {{formula:2e8d612a-e970-4c39-a56f-73def10cb1fc}} whose corresponding singular values are greater than a threshold. In section 6 we combine these components to obtain a quantum recommendation algorithm and analyze its performance and running time."
] |
[] |
Algorithms for Batch Hierarchical Reinforcement Learning
|
Hierarchical Reinforcement Learning (HRL) exploits temporal abstraction tosolve large Markov Decision Processes (MDP) and provide transferable subtaskpolicies. In this paper, we introduce an off-policy HRL algorithm: HierarchicalQ-value Iteration (HQI). We show that it is possible to effectively learnrecursive optimal policies for any valid hierarchical decomposition of theoriginal MDP, given a fixed dataset collected from a flat stochastic behavioralpolicy. We first formally prove the convergence of the algorithm for tabularMDP. Then our experiments on the Taxi domain show that HQI converges fasterthan a flat Q-value Iteration and enjoys easy state abstraction. Also, wedemonstrate that our algorithm is able to learn optimal policies for differenthierarchical structures from the same fixed dataset, which enables modelcomparison without recollecting data.
|
[
"There are three major approaches developed relatively independently {{cite:05ae402691dc18c0d8b308ae1e4f0c57b59d3d1e}}, aiming to formalize the idea of abstraction into reinforcement learning. The three approaches are: 1) the option framework {{cite:9a78a2d0ecf7f98b3c1bbf77a3bbc0575823a5ea}}, 2) Hierarchies of Abstract Machines (HAMs) {{cite:a4be5494774fdc1684c45ffd322d9edd3d948a71}} and 3) MAXQ framework {{cite:1037d3a2abb3a2c7dd64c8c76b0c3ebd0fe71d33}}.",
"Under the option framework, the developers augment the original action set by options, which are macro actions that have their own predefined policy, termination state and active state. Sutton et al have shown that such a system is a semi-Markov Decision Process (SMDP), which converges to a unique hierarchical optimal solution using a modified Q-learning algorithm.For HAM framework, rather than giving out the entire policy of these macro actions, developers just need to specify a partial program that specifies a part of the policy. Using HAMQ learning {{cite:a4be5494774fdc1684c45ffd322d9edd3d948a71}}, HAM can also converge to a hierarchical optimal solution.",
"At last, MAXQ framework provides an elegant formulation that decomposes the original MDP into several subroutines in a hierarchy and the algorithm can learn policies recursively for all the subroutines. Therefore, in the MAXQ framework, there is no need to specify the policy for any macro-actions. However, Dietterich shows that it can only achieve recursive optimal solution, which in the extreme case, can be arbitrarily worse than the hierarchical optimal solution.",
"All of the above work assumes that the agent can interact with the world while learning. However, in real-world applications that needs HRL, it is usually very expensive to collect data and terrible failures are not allowed on operation. This forbids the usage of online learning algorithms that could potentially preform horribly in the early learning stage. To our best knowledge, there is little prior work {{cite:9d7853cd4a3669dedcad58fe8c80d5d331ff4272}} in developing batch learning algorithms that allow a hierarchical SMDP to be trained from an existing data set collected from a stochastic behavior policy. We believe that such algorithms are valuable for applying HRL in complex practical domains."
] |
[] |
Trajectory Aligned Features For First Person Action Recognition
|
Egocentric videos are characterised by their ability to have the first personview. With the popularity of Google Glass and GoPro, use of egocentric videosis on the rise. Recognizing action of the wearer from egocentric videos is animportant problem. Unstructured movement of the camera due to natural headmotion of the wearer causes sharp changes in the visual field of the egocentriccamera causing many standard third person action recognition techniques toperform poorly on such videos. Objects present in the scene and hand gesturesof the wearer are the most important cues for first person action recognitionbut are difficult to segment and recognize in an egocentric video. We propose anovel representation of the first person actions derived from featuretrajectories. The features are simple to compute using standard point trackingand does not assume segmentation of hand/objects or recognizing object or handpose unlike in many previous approaches. We train a bag of words classifierwith the proposed features and report a performance improvement of more than11% on publicly available datasets. Although not designed for the particularcase, we show that our technique can also recognize wearer's actions when handsor objects are not visible.
|
[
"Action recognition has been a popular problem in computer vision. However, this is typically done from a third person view, for example, from a static or a handheld camera. A standard line of work is to encode the actions using keypoints and descriptors. This is done by extending spatial domain descriptors to space-time descriptors. These descriptors are then matched using Euclidean distance or other similar measures. Some techniques also rely on supervised learning with these descriptor vectors. Some notable contributions in this area includes stip {{cite:0c3e1dfdeab701603bd2f334f7fcd2ece8abe9d5}}, 3d-sift {{cite:14ea387b3329f4fa486b6ea9fc572db35ec7dd1c}}, hog3d {{cite:2aad0e7210bffc2e40f69adb48f1fdfb038044fe}}, extended surf {{cite:72bd7a5fc0c8e969f9fb25d3cd55cb12a612cfca}}, and Local Trinary Patterns {{cite:9331ceaf78747bc4c723878ecba2480b506e1536}}. Methods that follow the pipeline of keypoint detection followed by an action descriptor, usually work on a cuboidal video volume. They tend to merge the optical flow and the appearance information from the foreground and objects present in the scene. There have been proposals to demerge these two. Such attempts track feature points in a video and use these trajectories as cues for the action recognition. Some recent methods {{cite:1ac11f4f527db20e2a1be22aa96b0365ad336631}}, {{cite:d1c6d71a8bf77c868e4d52bf18bcfbb2902fb119}}, {{cite:f3567dc02db312eb662277e14caaa648d2d68108}}, {{cite:3cb3e92b500ff84dd2d35772d8c0c5e4edb93159}} show promising results for action recognition by leveraging the motion information of trajectories.",
"Camera motion is very common in real-world videos and poses a significant challenge to any action recognition technique. Wang {{cite:7f74eb8e5b827242519257e4c346e644e10caf93}} propose a descriptor based on motion boundaries, to reduce the interference from camera motion. They compute motion boundaries by a derivative operation on the optical flow field. Thus, motion due to locally translational camera movement is canceled out and relative motion is captured. There have been various improvisation of the technique {{cite:1e3e27d8854a709d8f72dd8372e84e3ca5d94252}}, {{cite:bec2174165445bd246d1f71205276a08482437f0}}, {{cite:b8de0351252d2814b1bf71018d38491d39c4349c}} decomposing visual motion into dominant and residual motions both for extracting trajectories and computing descriptors.",
"Egocentric cameras have certain distinct advantages as well as constraints for action recognition. While having much lesser occlusions for the objects in an egocentric video is extremely useful, natural head motion of the wearer brings in additional large camera motion, posing a challenge to any first person action recognition algorithm. Within egocentric community, Spriggs {{cite:81e7481b908db5b7760ad43c32a1798da37f034d}} proposed to recognize first person actions using a mixture of gist {{cite:6c0632f94fdd645353e8553b52f6ce07d14a028d}} features and imu data. Their results confirm the importance of head motion in the first person action recognition. Pirsiavash and Ramanan {{cite:67ec680a3524062d53f9fb539ce54c04c61c49fe}} attempt to recognize the activity of daily living (adl). Their thesis is that the first person action recognition is “all about the objects\", and in particular, “all about the objects being interacted with\". To recognise the objects from first person view, they develop representations including {{formula:bf078743-907a-4174-9dcf-636a2a507d2e}} temporal pyramids, which generalize the well-known spatial pyramid to approximate temporal correspondence when scoring a model and {{formula:76f34678-d94d-488b-ac47-fd14def0762a}} composite object models that exploit the fact that objects look different when being interacted with. McCandless and Graumann {{cite:776ef19f71b74ac7dd3b0ac2656546b63f0a4e3e}} extend the work by using spatio-temporal pyramid histograms of objects appearing in the action. They devise a boosting approach that automatically selects a small set of useful spatio-temporal pyramid histograms among a randomized pool of candidates. In order to efficiently focus on the candidates, they propose an “object-centric\" scheme that prefers candidates involving objects prominently involved in the actions. Fathi {{cite:a54d26c466b9a995a145fa1703e8c134edbe20b2}} recognize the importance of hands in the first person action recognition. They propose a representation for egocentric actions based on hand-object interactions and include cues such as optical flow, pose, size and location of hands in their feature vector. There is an assumption on the availability of hand, object and background labels in the video. Objects are not always the most important cue in the first person action recognition. In a sports video, when there are no prominent handled object, Kitani {{cite:c2c3fbccda7be889c633402ac2cd54eede8330d3}} use motion based histograms recovered from the optical flow of the scene (background) to recognize the actions of the wearer. Ogaki {{cite:b73b4c4954f105e0775d076b9e4ba146f44d055a}} use eye-motion and ego-motion to recognize indoor desktop actions. Recently, Ryoo have suggested pooled motion features tracking how descriptor values are changing over time and summarizing them to represent an action in the video {{cite:7fdc9a819fe9e40eb56ada2315124f9ff4f67022}}. In a parallel independent work, Li have also proposed a feature descriptor based upon dense trajectories {{cite:e6f307236cfd21210a8c7f1cb751a0a204030a45}}. However they also use complex patterns like gaze and hand pose which we show are not necessary to reach state of the art accuracy."
] |
[] |
Bridging the Gaps Between Residual Learning, Recurrent Neural Networks
and Visual Cortex
|
We discuss relations between Residual Networks (ResNet), Recurrent NeuralNetworks (RNNs) and the primate visual cortex. We begin with the observationthat a special type of shallow RNN is exactly equivalent to a very deep ResNetwith weight sharing among the layers. A direct implementation of such a RNN,although having orders of magnitude fewer parameters, leads to a performancesimilar to the corresponding ResNet. We propose 1) a generalization of both RNNand ResNet architectures and 2) the conjecture that a class of moderately deepRNNs is a biologically-plausible model of the ventral stream in visual cortex.We demonstrate the effectiveness of the architectures by testing them on theCIFAR-10 and ImageNet dataset.
|
[
"Deep Recurrent Neural Networks: Our final model is deep andsimilar to a stacked RNN{{cite:81be0a6bc2802d8e0ca310f7c51b4a70eaba9459}}, {{cite:01cd4f681d5cb9b8905e67160c92105745b8bd7a}}, {{cite:12a1676ea7647e02a5248b2ba82f1f7a9782a510}}with several main differences: 1. our model has feedback transitionsbetween hidden layers and self-transition from each hidden layer toitself. 2. our model has identity shortcut mappings inspired byresidual learning. 3. our transition functions are deep andconvolutional.",
"As suggested by {{cite:98d9b94e219a614b31580d1496a6a1b85ccd25e6}}, the term depth in RNNcould also refer to input-to-hidden, hidden-to-hidden orhidden-to-output connections. Our model is deep in all of thesesenses. See Section REF .",
"Recursive Neural Networks and Convolutional Recurrent NeuralNetworks: When unfolding RNN into a feedforward network, theweights of many layers are tied. This is reminiscent of RecursiveNeural Networks (Recursive NN), first proposed by{{cite:3dc3f2c4bba93e4b19444f392e333d3a19f7af9d}}. Recursive NN are characterized by applyingsame operations recursively on a structure. The convolutional versionwas first studied by {{cite:dae591d1d967de3916d2da8650f9ca792d76a80f}}. Subsequent relatedwork includes {{cite:23a5812c7b8dc55c674ec7adb9aefcf724207e68}} and{{cite:312cfadd6c36c42af03625433becab525dcae605}}. One characteristic distinguishes our modeland residual learning from Recursive NN and convolutional recurrent NNis whether there are identity shortcut mappings. This discrepancyseems to account for the superior performance of residual learning andof our model over the latters.",
"A recent report {{cite:5d93d9c4b8f4bd02b4a189f3b312c0e4ef9bc92f}} we became aware of after wefinished this work discusses the idea of imitating cortical feedbackby introducing loops into neural networks.",
"A Highway Network {{cite:5cf90ded4045ef037f3aee228d6d44bbcbeb05ab}} is a feedforward network inspired by Long Short Term Memory {{cite:6c325daf4fb7d3cae0f57a79be37303d358d751d}} featuring more general shortcut mappings (instead of hardwired identity mappings used by ResNet)."
] |
[] |
Almost All Even Yao-Yao Graphs Are Spanners
|
It is an open problem whether Yao-Yao graphs $\mathsf{YY}_k$ (also known assparse-Yao graphs) are all spanners when the integer parameter $k$ is largeenough. In this paper we show that, for any integer $k\geq 42$, the Yao-Yaograph $\mathsf{YY}_{2k}$ is a $t_k$-spanner, with stretch factor$t_k=6.03+O(k^{-1})$ when $k$ tends to infinity. Our result generalizes thebest known result which asserts that all $\mathsf{YY}_{6k}$ are spanners for$k$ large enough [Bauer and Damian, SODA'13]. Our proof is also somewhatsimpler.
|
[
"Replacing the Euclidean distance {{formula:23253064-2269-44d4-8242-8013c32c53e5}} by a power {{formula:87cad8fd-cdcf-418d-87db-3a90a1f91d5f}} with a constant {{formula:eef510cb-592f-4937-b21e-4c8fbb9757e5}} leads to the definition of power spanners.Since the power of length models the energy consumed in wireless transmissions,power spanners have important implications in wireless networking applications.In this setting, Yao-Yao graphs {{formula:aa3501c4-89f4-4df4-af5c-485d439d4ea3}} have been proved to be power spanners for any {{formula:64999382-0eb8-4029-91de-05f48bffbbe5}} when {{formula:d944fa3b-582f-4e94-a2b9-bbffb342d0f0}} {{cite:15906211f3872cbfdfb9a8b003f31c4ca8b0e04a}}, {{cite:1ea01c9e29656248b48a10d78a82d0c016896c0c}}.It is clear that when a graph is a geometric spanner, it must also be a power spanner (See {{cite:caf19ec24a5ea74576804c938e52bfeaecd77ad0}}),however the reverse does not hold.",
"We also note that neither Yao graphs nor Yao-Yao graphs can be guaranteed to be planar graphs{{cite:1ea01c9e29656248b48a10d78a82d0c016896c0c}}, {{cite:a3758a68605bc82b47648361163abbb3357940d1}},whereas Delaunay triangulation provides another type of spanner which is planar but without bounded degree{{cite:0b573bde31702ea9e1b5e604a968ab718112fc1d}}, {{cite:2d4131e8904b386f4246104c50a7babe47cddf04}}.In order to facilitate network design in certain applications, some previous work {{cite:be105045bb598de8d994a26e5ab3e43ff3f85aad}}, {{cite:fd86ff7260710aebb658e4efd919f9a10cea7ed0}}made use of bothYao graphs and Delaunay graphs to produce degree-bounded and planar geometric spanners.",
"The paper is organized as follows. We introduce some standard notations and useful tools,including the shape of curved trapezoid, in Section .We introduce overlapping Yao graphs, trapezoidal Yao graphs, and prove their spanning propertiesin Section . We proveour main result in Section .Finally, we conclude with some future work in Section ."
] |
[] |
Accurate Online Full Charge Capacity Modeling of Smartphone Batteries
|
Full charge capacity (FCC) refers to the amount of energy a battery can hold.It is the fundamental property of smartphone batteries that diminishes as thebattery ages and is charged/discharged. We investigate the behavior ofsmartphone batteries while charging and demonstrate that the battery voltageand charging rate information can together characterize the FCC of a battery.We propose a new method for accurately estimating FCC without exposinglow-level system details or introducing new hardware or system modules. We alsopropose and implement a collaborative FCC estimation technique that builds oncrowdsourced battery data. The method finds the reference voltage curve andcharging rate of a particular smartphone model from the data and then comparesthe curve and rate of an individual user with the model reference curve. Afteranalyzing a large data set, we report that 55% of all devices and at least onedevice in 330 out of 357 unique device models lost some of their FCC. For somemodels, the median capacity loss exceeded 20% with the inter-quartile rangebeing over 20 pp. The models enable debugging the performance of smartphonebatteries, more accurate power modeling, and energy-aware system or applicationoptimization.
|
[
"Significant amount of research work focused on the energy consumption measurement and optimization of different applications and system {{cite:3deaade2c75a0d34e2f207d8b52750cbe452e25d}}, {{cite:4608e592b60ca41fd2c1606055ccf69b61febae1}}, {{cite:d026966abdd87b215910d57f681413a10b3c5c2b}}, {{cite:60511bd0c31934e9055caf709ecb205720d8dfd2}}, {{cite:64afe2bb5845fc73910ff3b3b9359783d2d79acd}}. A large body of research investigated the energy consumption of mobile devices through profiling, modeling, and debugging {{cite:d43fa81725ef6c271fedab6db5d2839c7e067eb2}}. The profiling methods include novel techniques to trace the energy consumption from code to different hardware components {{cite:a8d3a5cddce0355a4f9e5466b64f3a5ac0d82e87}}, {{cite:660ed2ff330563181e1371824052b5d63dd6695e}}, {{cite:60511bd0c31934e9055caf709ecb205720d8dfd2}}. Such profilers also depend on the power consumption modeling. The state-of-the art power modeling approaches rely on on-device resource profiling and battery information. PowerBooter {{cite:00a9097c876e9886fe5519cbca7d473be7e965e9}} relies on SOC updates and OCV discharge curves to build the regression based power models. V-edge {{cite:c49254d490d3127a5b6d8a47452cbce66f0ce4e0}} and BattTracker {{cite:98cbe40f5cc11082d1d2d978df3de1726f67a141}} rely on SOC updates, OCV and the load voltage to model power consumption. Sesame {{cite:a8d3a5cddce0355a4f9e5466b64f3a5ac0d82e87}}, AppScope {{cite:425767f8fedfc834330250139cb82c5d675660dc}}, and DevScope {{cite:fecec0e3cca8c601e93e076ad31e5311cf31c7a1}} rely on SOC updates and current drawn estimates from the Coulomb counter-based fuel gauges. Carat relies on SOC updates to compute the energy ratings of different applications for different energy models towards finding the energy hogs and bugs. We also depend on battery manager updates. We specifically use SOC update time to find the charging rates and use the battery voltage to determine the length of the CC phase. However, we estimate the FCC of the battery. Our approach works irrespective the SOC estimation or fuel gauge chip used by the device.",
"The discharge rate of the battery increases for the same usage {{cite:926ef72bbe24db9d2dcc6a3e72e330c674110382}}, as the Lithium-Ion battery ages. Consequently, the reported power consumption will be higher than the true value for an aged battery. The accuracy of the self-metering profilers is intertwined with the age of the battery and will decrease as the age of the battery increases.For this reason, the self-metering profilers may require retraining of their models. Precise quantification of its effect on the accuracy would require further evaluation of the profilers with batteries of different ages and so far has been an open question. Battery FCC estimation model enables more accurate modeling by identifying the inefficiency of the energy source of the smartphone.",
"A number of data-driven strategies exist that predict the capacity as the battery ages and the number of charging cycle increases. Yin et al. {{cite:78e5535b54b5d75777f35cfa3ae5367457f3bbc4}} and Liu et al. {{cite:bf513ac54c36c9f5393268cacf3d4292e7706a5c}} applied a few variations of the Gaussian process regression to predict the capacity as a function of charging cycle. The earlier approaches rely on the vendor provided data on FCC and charging cycles. Barré et al. {{cite:926ef72bbe24db9d2dcc6a3e72e330c674110382}} studied how the discharge rates can be used to predict the life of Lithium-Ion battery in electric vehicles. Based on the discharge voltage and current during an activity, such as acceleration, they proposed a data driven real-time battery capacity prediction framework.",
"Compared to these related work, we focus on estimating capacity proactively. Our technique does not require any additional hardware or system modification. Therefore, the approach can be easily implemented as a part of the mobile system and can be integrated with different applications or operating system initiated optimizations. For example, the SDB proposed by Badam et al. {{cite:00d4a304db91171de21e27c31dbe670b39c5144b}} can include a FCC aware battery scheduling, where SDB will learn the FCC of an individual battery while charging and then schedule the batteries accordingly while discharging."
] |
[] |
A Novel Approach to Dropped Pronoun Translation
|
Dropped Pronouns (DP) in which pronouns are frequently dropped in the sourcelanguage but should be retained in the target language are challenge in machinetranslation. In response to this problem, we propose a semi-supervised approachto recall possibly missing pronouns in the translation. Firstly, we buildtraining data for DP generation in which the DPs are automatically labelledaccording to the alignment information from a parallel corpus. Secondly, webuild a deep learning-based DP generator for input sentences in decoding whenno corresponding references exist. More specifically, the generation istwo-phase: (1) DP position detection, which is modeled as a sequentiallabelling task with recurrent neural networks; and (2) DP prediction, whichemploys a multilayer perceptron with rich features. Finally, we integrate theabove outputs into our translation system to recall missing pronouns by bothextracting rules from the DP-labelled training data and translating theDP-generated input sentences. Experimental results show that our approachachieves a significant improvement of 1.58 BLEU points in translationperformance with 66% F-score for DP generation accuracy.
|
[
"There is some work related to DP generation. One is zero pronoun resolution (ZP), which is a sub-direction of co-reference resolution (CR). The difference to our task is that ZP contains three steps (namely ZP detection, anaphoricity determination and co-reference link) whereas DP generation only contains the first two steps. Some researchers {{cite:cd0c3b18f7266e9f1d38b486bd5f5bf5db1b677b}}, {{cite:4663393982875b59c08d2619a598c0d55e26d3b2}}, {{cite:7f626cd03b93c75ee2ed9066e7b70790aeb1247d}} propose rich features based on different machine-learning methods. For example, Chen:2013:EMNLP propose an SVM classifier using 32 features including lexical, syntax and grammatical roles etc., which are very useful in the ZP task. However, most of their experiments are conducted on a small-scale corpus (i.e. OntoNotes)It contains 144K coreference instances, but only 15% of them are dropped subjects. and performance drops correspondingly when using a system-parse tree compared to the gold standard one. Novak:2014:COLING explore cross-language differences in pronoun behavior to affect the CR results. The experiment shows that bilingual feature sets are helpful to CR. Another line related to DP generation is using a wider range of empty categories (EC) {{cite:ad2315322f18a28972844bad4d65fa62606a6edf}}, {{cite:2ed2018a03c58cf522ccbd2feec792da66ab867b}}, {{cite:8b93d1968f07e3128d65a6f3e6e94273f6b7d4e8}}, which aims to recover long-distance dependencies, discontinuous constituents and certain dropped elementsEC includes trace markers, dropped pronoun, big PRO etc, while we focus only on dropped pronoun. in phrase structure treebanks {{cite:2b05b137e174254c539ea1e0bd0a9df1dab9ea0f}}. This work mainly focus on sentence-internal characteristics as opposed to contextual information at the discourse level. More recently, yang2015recovering explore DP recovery for Chinese text messages based on both lines of work.",
"These methods can also be used for DP translation using SMT {{cite:c474303fc6d5357976815a377491b2149526c56a}}, {{cite:e89460e2ebedeb2e83041a906f72545eb8490e9a}}, {{cite:cdd348622a1ba281153ceadf07ac65e9f0a3476b}}, {{cite:1ecd875cfabe087ea8272fb9b6497483f4566e23}}. Taira:2012:SSSST propose both simple rule-based and manual methods to add zero pronouns in the source side for Japanese–English translation. However, the BLEU scores of both systems are nearly identical, which indicates that only considering the source side and forcing the insertion of pronouns may be less principled than tackling the problem head on by integrating them into the SMT system itself. Nagard:2010:ACL present a method to aid English pronoun translation into French for SMT by integrating CR. Unfortunately, their results are not convincing due to the poor performance of the CR method {{cite:d29160b9dd80af6620da2e7d7f22f783f0812ef7}}. chung2010effects systematically examine the effects of EC on MT with three methods: pattern, CRF (which achieves best results) and parsing. The results show that this work can really improve the end translation even though the automatic prediction of EC is not highly accurate."
] |
[] |
Tree-Deletion Pruning in Label-Correcting Algorithms for the
Multiobjective Shortest Path Problem
|
In this paper, we re-evaluate the basic strategies for label correctingalgorithms for the multiobjective shortest path (MOSP) problem, i.e., node andlabel selection. In contrast to common believe, we show that---when carefullyimplemented---the node-selection strategy usually beats the label-selectionstrategy. Moreover, we present a new pruning method which is easy to implementand performs very well on real-world road networks. In this study, we test ourhypotheses on artificial MOSP instances from the literature with up to 15objectives and real-world road networks with up to almost 160,000 nodes.
|
[
"The techniques used for solving the MOSP problem are based on labeling algorithms.The majority of the literature is concerned with the biobjective case.The latest computational study for more than 2 objectives is from 2009 {{cite:48c2e646df81c1129f6848f4c3f81baf892e1050}} and compares 27 variants of labeling algorithms on 9,050 artificial instances.These are the instances we also use for our study.In summary, a label correcting version with a label-selection strategy in a FIFO manner is concluded to be the fastest strategy on the instance classes provided.The authors do not investigate a node-selection strategy with the argument that it is harder to implement and is less efficient (cf. also {{cite:84ae486d15bd1a56ab257cb5ba989ae51b20e2ea}}).",
"In an older study from 2001 {{cite:ed4ff5883ae598adfd7d14d7ec2d82e013b55d91}}, also label-selection and node-selection strategies are compared.The authors conclude that, in general, label-selection methods are faster than node-selection methods.However, the test set is rather small, consisting of only 8 artificial grid-graph instances ranging from 100 to 500 nodes and 2 to 4 objectives and 18 artificial random-graph instances ranging from 500 to 40,000 nodes and densities of 1.5 to 30 with 2 to 4 objectives.",
"In the work by Delling and Wagner {{cite:4b5f05252bee915286a3c583fa3d4b461ffb0071}}, the authors solve a variant of the multiobjective shortest path problem where a preprocessing is allowed and we want to query the Pareto-front of paths between a pair of nodes as fast as possible.The authors use a variant of SHARC to solve this problem.Though being a different problem, this study is the first computational study where an implementation is tested on real-world road networks instead of artificial instances.The instances have sizes of 30,661, 71,619 and 892,392 nodes and 2 to 4 objectives.Though, the largest instance could only be solved using highly correlated objective functions resulting in Pareto-front sizes of only at most 2.5 points on average."
] |
[] |
Finding Influential Users in Social Media Using Association Rule
Learning
|
Influential users play an important role in online social networks sinceusers tend to have an impact on one other. Therefore, the proposed workanalyzes users and their behavior in order to identify influential users andpredict user participation. Normally, the success of a social media site isdependent on the activity level of the participating users. For both onlinesocial networking sites and individual users, it is of interest to find out ifa topic will be interesting or not. In this article, we propose associationlearning to detect relationships between users. In order to verify thefindings, several experiments were executed based on social network analysis,in which the most influential users identified from association rule learningwere compared to the results from Degree Centrality and Page Rank Centrality.The results clearly indicate that it is possible to identify the mostinfluential users using association rule learning. In addition, the resultsalso indicate a lower execution time compared to state-of-the-art methods.
|
[
"Online social networks and social media analysis are popular research areas in contemporary network science. The main focus in social network research is on link prediction {{cite:a9b07ad58af4e2070d05d569c8fd18bd7310a8c8}} and social connection prediction {{cite:f4d3813425c77291c488bbe2ce5c6e217b0cc2a1}}. Different teams around the world also work on: (i) personality prediction for micro blog users {{cite:ccc20b3a45900ac95c78759cebc5e8d99bccf3aa}}, (ii) churn prediction and its influence on the network {{cite:afd31097e2712b861966d86f3e01376ac22ddaa3}}, {{cite:450c579ff12288cc8931ad006847d51a224f88cc}}, (iii) community evolution prediction {{cite:179c3f3407b9a9d860a3b0e0f9e7cd15f427a34f}}, {{cite:e5689184b8d8735ac41c8e6c2584986ebdbabe60}}, (iv) using social media to predict real-world outcomes {{cite:7db2258148795939f2d8446335e0e4c596fe3caa}}, (v) predicting friendship intensity {{cite:60b67dfa02023efce9ee4e66a23238491746703a}}, {{cite:9d038a147a8a03d0c8ce41922e8c2eb96c023c0b}}, (vi) affiliation recommendations{{cite:a9b81608484c2a1d3f1fa816e8877d1926002b89}}, {{cite:c78c676979ac56020ddd02c0c5a2f4061b84c1ed}}, and (vii) sentiment analysis and opinion mining {{cite:7c7dfb4a5c5f1bf600deb993638f0be1bd2451ce}} .",
"Other popular areas of research focus on popularity prediction in social media based on comment mining {{cite:74e38821b163414c548cbd62d62aa74b291a050b}}, predicting information cascade on social media {{cite:f98f1a99ba49b9242373e25744410d48d3995a4a}}, and predicting patterns of diffusion processes in social network {{cite:29d21afd520d1e2344f05b71cd8fe293adceef12}}. An important factor is often the user's role in the different processes. As such, identifying influential users are of interest to understand and/or affect the spread of information, e.g., viral marketing. The ability to identify influential users might also affect the research into other areas of related work (e.g., ii or iii).",
"Research into detecting influential users on Twitter indicates that, while a large amount of followers seem to be present among influential users, predictions of which particular user will be influential is unreliable {{cite:cd02da4c05d90cded4e4ae8968e158c98fc2184e}}. Depending on the social network, how to define influence differs, e.g., influence on Twitter might be defined by retweets or mentions, while, on Digg, votes generated are used to measure influence {{cite:b12dad7b2e48c0dda29703c3e5e43dc340df6f5c}}, {{cite:5deb6121c69e15a7cc57678138ae1e6f973cf581}}, {{cite:58b074afcdbf57a0b67f8f9756bc241176fb9c31}}. While some initial research has been done using clustering algorithms to identify top users, based on influence features, e.g., likes and replies, evaluation is lacking {{cite:0e027c9c183ece057f856223e7eca9dc6c3bd808}}. Similarly, linear regression has been used to identify influential (categorical) users based on influence features {{cite:58b074afcdbf57a0b67f8f9756bc241176fb9c31}}.",
"While some research on identifying influential users use learning based approaches, another popular approach to identifying influential users is the Page Rank algorithm or adaptions of the Page Rank algorithm {{cite:63f91de3b717be38918a46d04d2c3103a39d8151}}, {{cite:1c6a913086946bc8720eaa1e1ed6c2eb7acbc36a}}, {{cite:c44730f161bab4e2bdda4345953bee67a957ebfe}}.",
"{{cite:91d84f7793d92d5b41dc4b7b4dbaaaea0f39adc4}} explore the association rule between a course and gender in the Facebook 100 university dataset. This was performed to discover the influence of gender in studying a specific course. {{cite:b2dce80818f017d927f3c26cda12642776246b2b}} introduce the scheme for association rule learning of personal hobbies in social networks, while {{cite:2fde63eb5132de39ac4f055f17cc789901fc1e35}} tackle the problem of mining association rules in folksonomies and try to find out how association rule learning can be applied to analyze and structure folksonomies.",
"Initial research used association rule learning to identify influential users and predict user participation in online social networks {{cite:663450317bd56e38e6e788acf700a548e3ef5fbc}}. Association rule learning has been previously used in social network and social media analysis.",
"While online social network analysis is popular, there is, according to our review, a lack of research on using association rules for predicting user participation in online social media discussions."
] |
[] |
Detecting Ground Control Points via Convolutional Neural Network for
Stereo Matching
|
In this paper, we present a novel approach to detect ground control points(GCPs) for stereo matching problem. First of all, we train a convolutionalneural network (CNN) on a large stereo set, and compute the matching confidenceof each pixel by using the trained CNN model. Secondly, we present a groundcontrol points selection scheme according to the maximum matching confidence ofeach pixel. Finally, the selected GCPs are used to refine the matching costs,and we apply the new matching costs to perform optimization with semi-globalmatching algorithm for improving the final disparity maps. We evaluate ourapproach on the KITTI 2012 stereo benchmark dataset. Our experiments show thatthe proposed approach significantly improves the accuracy of disparity maps.
|
[
"Surveys regarding stereo methods we refer readers to the taxonomy of Scharstein and Szeliski{{cite:84fbb68c933212ee6786dabafec56de382ca6246}} and its companion website.",
"In this section, we first briefly review the learning-based methods for detecting ground control points, or predicting the confidence of matching costs. Then, the discussion of several stereo matching costs computation methods based on Convolutional Neural Network.",
"An early learning-based approach to stereo matching proposed by Lew et al.{{cite:75a81ccedc648b9caafdbe0dea4c9000d5bd7500}} adopted instance based learning (IBL) to select optimal feature set points for stereo matching.Kong and Tao {{cite:2be55d74df603ce981ad37913a05f4ae41b4ffab}} used nonparametric techniques to train a model to predict the probability of a potential match over three categories: correct, incorrect due to foreground over-extension, and incorrect due to other reasons. The predicted knowledge was then integrated into an MRF framework to improve the depth computation.Later on, Kong and Tao{{cite:373147fad87e3b55af0154b6dbc1387d1249ec48}} extended their work by learning multiple experts from different normalized cross-correlation (NCC) matching windows sizes and centers, and the likelihood under each expert was then combined probabilistically into a global MRF framework for improving accuracy.Motten et al. {{cite:cbb8237f09d77a4c42348e56fa3e3cd1b8283674}} trained a hierarchical classifier for selecting the most promising disparity with the matching costs and spatial relationship of pixels.Peris et al. {{cite:940a7c9172cf795a7b5c45abc9528385003ee6fa}} designed a feature from cost volume, and computed the final stereo disparity using Multiclass Linear Discriminant Analysis (Multiclass LDA).",
"More recently, in {{cite:24be00beb7d00f4fa44faafbbfa87d56db873311}}, {{cite:7b9eba0047f1c2b2777abdb1ce06d817dbbf173e}}, {{cite:a9948ef9ea486198740e9a4f34508cb1a8356279}}, they employed random decision forests to estimate the confidence of the stereo matching costs.Haeusler et al. {{cite:24be00beb7d00f4fa44faafbbfa87d56db873311}} trained a random decision forests classifier to predict the confidence of stereo matching costs by combining several confidence measures into a feature vector.Similarity, Spyropoulos and Modorhai {{cite:a9948ef9ea486198740e9a4f34508cb1a8356279}} used a random decision forests classifier to estimate the confidence of the matching costs, and showed that the confidence information can further be used in a Markov random field framework for improving stereo matching.Park and Yoon {{cite:7b9eba0047f1c2b2777abdb1ce06d817dbbf173e}} exhibited that effective confidence measures can be selected by estimating the permutation importance of each measure. In addition, they applied the confidence value to modify the initial matching cost, and inserted the new matching costs into stereo method to decrease the error of stereo matching.All of these learning-based approaches have to structure a set features of confidence measures, and train a classifier based on specified matching costs, which indicates that different classifiers should be trained when applying different stereo matching cost computation methods.",
"Convolutional Neural Networks (CNN) has been rapidly developed in recent years, and has beenextensively applied to deal with various computer vision tasks.The most recent year, CNN has been used in stereo matching {{cite:e268e9a8ff006c95304b68ec487cc8d9ef9e9571}}, {{cite:e2ca86701a46bb2d84291d2528c7382761bb2ea2}}, {{cite:e9bc3307ac2c81ecae67febe05ea94b4e8a93c03}}, {{cite:79d78ced11b39f884939feb0700f868ed15d3a4b}} and achieved noticeable results. The destination of these methods is training a CNN with a large set of stereo image patches, and comparing the matching cost between image patches by the trained network. The main difference between them is the architecture of the network. On the Contrast to these methods, we train a CNN to compute matching confidence of each pixel, and detect the GCPs based on the output of CNN.{{figure:aa82399e-f694-42ff-8022-d35021bef993}}"
] |
[] |
Predicting Performance on MOOC Assessments using Multi-Regression Models
|
The past few years has seen the rapid growth of data min- ing approaches forthe analysis of data obtained from Mas- sive Open Online Courses (MOOCs). Theobjectives of this study are to develop approaches to predict the scores a stu-dent may achieve on a given grade-related assessment based on information,considered as prior performance or prior ac- tivity in the course. We develop apersonalized linear mul- tiple regression (PLMR) model to predict the grade fora student, prior to attempting the assessment activity. The developed model isreal-time and tracks the participation of a student within a MOOC (viaclick-stream server logs) and predicts the performance of a student on the nextas- sessment within the course offering. We perform a com- prehensive set ofexperiments on data obtained from three openEdX MOOCs via a Stanford Universityinitiative. Our experimental results show the promise of the proposed ap-proach in comparison to baseline approaches and also helps in identification ofkey features that are associated with the study habits and learning behaviorsof students.
|
[
"Several researchers have focused on theanalysis of education data (including MOOCs), inan effort tounderstand the characteristics ofstudent learning behaviors and motivation within this education model {{cite:1ae62e17d64f6b1e492479ed7e0a4367a894ac8a}}. Boyer et. al. {{cite:0aa61f9519133d374b5be3a1eb5a90691f6e0683}} focus onthe stopout prediction problem within MOOCs; bydesigning a set of processes using information fromprevious courses and the previous weeks of thecurrent course.Brinton et. al. {{cite:668e6d570b92b580812d92a5d80d60d29c714734}} developed an approach topredict if a student answers a question correct on thefirst attempt via click-stream information and social learningnetworks. Kennedy et. al. {{cite:f1a297d0e194832004136eaf9b75738fee404b71}}analyzed the relationship between a student'sprior knowledge on end-of-MOOC performance.Sunar et. al. {{cite:a424efe0b2c004f359570e8319a3f79526b31211}} developed an approachto predict the possible interactions between peers participating in a MOOC.",
"Most similar to our proposed work,Bayesian Knowledge Tracing (BKT) {{cite:d7f8eb7473ddac3bb2b2f3c5a17ff04f6fd4aa0c}} has been adapted topredict whether a student can get a MOOC assessment correct or not.BKT was first developed {{cite:e5c1cd5ea430326eab1b7a5e57869b64d8fe79c6}} for modeling theevolving knowledge states of students monitored within IntelligentTutoring Systems (ITS). Pardos et. al.proposed a model Item Difficulty Effect Model (IDEM)that incorporates the difficulty levelsof different questions and modifies the original BKT byadding an Item node to every question node. By identifyingthe challenges associated withmodeling MOOC data, the IDEM approach and extensions that involvesplitting questions into several sub-parts and incorporating resource (knowledge) information {{cite:c86e92b184949b3f2b1945b1afc5a4dc9097f269}}are considered state-of-the-art MOOC assessmentpredictionapproaches and referred as KT-IDEM.However, this approach can only predict a binary value grade. In contrast, the model proposed in this paper isable to predict both, a continuous and a binary grade.",
"Within learning analytics literature, outside of MOOC analysis,predicting student performance is a popular and extensive topic.Wang et. al. {{cite:719dec1daa0256a1dffea2445b0b04a44e82ec23}} performed a study to predict student's performance by capturingdata relevant to study habits and learning behaviorsfrom their smartphones. Specific examples of data captured includelocation, time, ambient noise and social activity. Coupled withself-reported information, this work captured the influence of a student'sdaily activity on the academic performance.Elbadrawy et. al. {{cite:5bbd400a240c1b8779d3e6b662c47f89b6fd5c9d}} proposed the use ofpersonalized linear multi-regression models to predict studentperformance in a traditional university by extracting data fromcourse management systems (Moodle). With a particularmembership vector for each student, the model was able to capturepersonal learning behaviorsand outperformed several baseline approaches. Ourstudy focuses on MOOCs, which presentsdifferent assumptions, challenges and features in comparison to a traditional universityenvironment."
] |
[] |
Machine Comprehension Based on Learning to Rank
|
Machine comprehension plays an essential role in NLP and has been widelyexplored with dataset like MCTest. However, this dataset is too simple and toosmall for learning true reasoning abilities. \cite{hermann2015teaching}therefore release a large scale news article dataset and propose a deep LSTMreader system for machine comprehension. However, the training process isexpensive. We therefore try feature-engineered approach with semantics on thenew dataset to see how traditional machine learning technique and semantics canhelp with machine comprehension. Meanwhile, our proposed L2R reader systemachieves good performance with efficiency and less training data.
|
[
"Machine comprehension generally concentrates on MCTest {{cite:e37cff713a33b3ad7a3f5975f828eb5ab4165975}} and due to the limitation of data size, the state of the arts are mainly based on traditional machine learning techniques. For example, {{cite:6491f821900b25eb7dbaa17b67a3a7ed7cbf3358}} proposed a max-margin learning framework that combines features on syntax, coreference, frame semantics and word embeddings, which achieves significant improvement on the problem of MCTest question answering. Although recently {{cite:ecde1a2e7d5158d98cf520ab29388584d96f4b14}} proposed a parallel-hierarchical model based on neural network and this method outperforms the previous feature-engineered approaches, it has reasoning limitation that the reasoning can only be achieved by stringing important sentences together. Their experiment proves that MCTest is too simple to learn true reasoning and it is also too small for that goal.",
"Considering the limitations of MCTest dataset, {{cite:14e1821f6c097d7ba81492890670d174bdf87f4f}} provides a large scale supervised reading comprehension dataset collected from the CNN and Daily Mail websites. This helps with the bottleneck that large dataset is missing on machine comprehension evaluation. With this dataset, {{cite:14e1821f6c097d7ba81492890670d174bdf87f4f}} propose a deep LSTM reader that achieves an accuracy of {{formula:cc0c94b3-ace7-4830-9e05-de4bcc0d3b8c}} on CNN and {{formula:5bec81d9-7c62-4e96-81f9-a80950bd96f8}} on Daily Mail. However, this deep LSTM reader is time consuming for training and no explanation can be found on why it works opposed to traditional approach. Therefore, we propose a traditional machine learning method on this new dataset to investigate what features can help with this task."
] |
[] |
Profit-Driven Team Grouping in Social Networks
|
In this paper, we investigate the profit-driven team grouping problem insocial networks. We consider a setting in which people possess differentskills, and the compatibility between these individuals is captured by a socialnetwork. Moreover, there is a collection of tasks, where each task requires aspecific set of skills and yields a profit upon completion. Individuals maycollaborate with each other as \emph{teams} to accomplish a set of tasks. Weaim to find a group of teams to maximize the total profit of the tasks thatthey can complete. Any feasible grouping must satisfy the following conditions:(i) each team possesses all the skills required by the task assigned to it,(ii) individuals belonging to the same team are socially compatible, and (iii)no individual is overloaded. We refer to this as the \textsc{TeamGrouping}problem. We analyze the computational complexity of this problem and thenpropose a linear program-based approximation algorithm to address it and itsvariants. Although we focus on team grouping, our results apply to a broadrange of optimization problems that can be formulated as cover decompositionproblems.
|
[
"To the best of our knowledge we are the first to formulateand study the team grouping problem and its variants. However, our work is closelyrelated to other team formation and cluster hiring problems. {{cite:3417dd633861ce4daf64c6ef80daccf052a90220}} introduce the minimum cost team formation problem. Given a set of skills that need to be covered and social network, their objective is to select a team of experts that can cover all required skills, while ensuring efficient communication between the team's members. There is a considerable amount of literature on this topic and its variants {{cite:a0768b1a036b43feb9f1965f58038fa7e24195b2}}, {{cite:9f6f43cbe97e87120dcb66a2992ebd54d405e2af}}, {{cite:94a64424514bd909774192d65ff21116bf7c84d5}}, {{cite:eaff57ab741ec1c27c752d2be2066303472e1a93}}, {{cite:75bed92d70ec0d71bb457e79f0de89bb152e23d7}}, {{cite:90dae96fff116e2753fe4e69e43a7e43257a2d1a}}. In {{cite:8e2b70f5c8ad6b1d8ecadddeb089967de66e0c42}}, they study cluster hire problem, where the objective is to hire a profit-maximizing team of experts with the ability to complete multiple projects, subject to a fixed budget. Different from all the above works where they aim to select a best qualified team, our objective is to group individuals into multiple teams. It turns out that these two problems are closely related, this allows us to leverage existing techniques for team formation to solve our problem.",
"The other category of related work is maximum disjoint set cover problem {{cite:66a05ac2e3feb100266c55a127cbaf811ec6a570}}. Given a universe, and a set of subsets, the objective is to find as many set covers as possible such that all set covers are pairwise disjoint. Our problem can be considered as a generalized disjoint set cover problem in the sense that every task in our problem may have different requirement of coverage, capacity constraint, and profit, and any feasible set cover must satisfy both coverage requirement as well as social compatibility. In addition, the requirement of “disjoint” is also relaxed by allowing individuals to have different load limits in our problem. Therefore, this work contributes fundamentally to the generalized cover decomposition problem."
] |
[] |
Robust and Efficient Relative Pose with a Multi-camera System for
Autonomous Vehicle in Highly Dynamic Environments
|
This paper studies the relative pose problem for autonomous vehicle drivingin highly dynamic and possibly cluttered environments. This is a challengingscenario due to the existence of multiple, large, and independently movingobjects in the environment, which often leads to excessive portion of outliersand results in erroneous motion estimation. Existing algorithms cannot copewith such situations well. This paper proposes a new algorithm for relativepose using a multi-camera system with multiple non-overlapping individualcameras. The method works robustly even when the numbers of outliers areoverwhelming. By exploiting specific prior knowledge of driving scene we havedeveloped an efficient 4-point algorithm for multi-camera relative pose, whichadmits analytic solutions by solving a polynomial root-finding equation, andruns extremely fast (at about 0.5$u$s per root). When the solver is used incombination with RANSAC, we are able to quickly prune unpromising hypotheses,significantly improve the chance of finding inliers. Experiments on syntheticdata have validated the performance of the proposed algorithm. Tests on realdata further confirm the method's practical relevance.
|
[
"The idea of deriving minimal case solvers for multi-camera relative pose is not new. In {{cite:fd0e48ae269723a17af5261b227d06d559faf7ae}}, Stewenius et al. proposed a minimal solution for the generalised epipolar equation using Groebner basis technique {{cite:95733f0e0ba414e7713a9970666462505a1f25d4}}. Despite only minimally 6 points are used for estimating the 6-DOF motion, there are totally 64 solutions to choose from, rendering RANSAC inefficient. Li et al. {{cite:05dc892538e00de16e447cc2a778285b3e6d4528}} proposed several linear solvers for generalised camera relative pose, among which the most general case is the 17-points algorithm. However, the large number of feature points needed prevent it from being used with RANSAC. Recently, Kneip and Li {{cite:c32d01b1dabba896a23ad00cf73cfc78f3007567}} presented an iterative solution based on eigenvalue minimisation. Although the objective function is in a closed form, its solution is sought via local iteration hence no converge guarantee. Ventura et al.{{cite:3fe4dfec0ff8d24eea11972016d649b1568146c8}} proposed to use first-order approximation of the camera rotation, by which the problem is reduced to solving a 20-th order polynomial. Numerically, solving a 20-th degree polynomial is sensitive to noise. Using one known directional correspondence to simplify motion estimation is not uncommon. Paper {{cite:e60bffcbe1d60d1b9764129bf3a6fbf658442525}}, proposed two efficient methods to solve the motion of a monocular camera. They give closed form solution to 4-order polynomial equation by utilising the fact that scale cannot be recovered in monocular vision.",
"Mostly close related to our method is {{cite:890891db4fb15fa7c005b9574b4c2228c6dbcd27}} in which the authors used IMU sensor to get two rotation readings and to solve the unknown yaw angles by solving an 8-th degree polynomial; our experiments however shown that the method is sensitive to image noise and the reference direction noise. In {{cite:c7a020105639ead9001ff1daed990ad8e8ceb056}}, Sweeney et al. proposed the start-of-the-art method to solve the problem by artfully select the angle-axis representation of rotation matrix, and this results the maximum 6 real solutions. A drawback is that when all Plücker lines correspondences are from the same cameras in the first and second frames, numerical unstable problem happens when the rotation is small which is the usual case in real-world. We do not compare methods using Ackerman vehicle motion model, as this model is restrictive in practice, and a post-relaxation is often needed (c.f. {{cite:2b8cf701b02c41e94911f85d83f3a066ffcadf9a}} {{cite:890891db4fb15fa7c005b9574b4c2228c6dbcd27}})."
] |
[] |
Joint Learning of Sentence Embeddings for Relevance and Entailment
|
We consider the problem of Recognizing Textual Entailment within anInformation Retrieval context, where we must simultaneously determine therelevancy as well as degree of entailment for individual pieces of evidence todetermine a yes/no answer to a binary natural language question. We compare several variants of neural networks for sentence embeddings in asetting of decision-making based on evidence of varying relevance. We propose abasic model to integrate evidence for entailment, show that joint training ofthe sentence embeddings to model relevance and entailment is feasible even withno explicit per-evidence supervision, and show the importance of evaluatingstrong baselines. We also demonstrate the benefit of carrying over textcomprehension model trained on an unrelated task for our small datasets. Our research is motivated primarily by a new open dataset we introduce,consisting of binary questions and news-based evidence snippets. We also applythe proposed relevance-entailment model on a similar task of rankingmultiple-choice test answers, evaluating it on a preliminary dataset of schooltest questions as well as the standard MCTest dataset, where we improve theneural model state-of-art.
|
[
"Our primary concern when integrating natural language query withtextual evidence is to find sentence-level representations suitableboth for relevance weighing and answer prediction.",
"Sentence-level representations in the retrieval + inference context have beenpopularly proposed within the Memory Network framework {{cite:ea2780af3c80e635aa7464401c3ddf2af3174d9c}},but explored just in the form of averaged word embeddings; the task includesonly very simple sentences and a small vocabulary.Much more realistic setting is introduced in the Answer Sentence Selectioncontext {{cite:5ff4f69a950da69a5e8d6752261c9379f85736b3}} {{cite:2a5dcfcad3c252fdfa9c1d5ad25d929deb1b2a80}}, with state-of-art models using complexdeep neural architectures with attention {{cite:aac51731a5b410c79c2a7c16234c929687c622bf}}, but the selectiontask consists of only retrieval and no inference (answer prediction).A more indirect retrieval task regarding news summarization was investigatedby {{cite:33826ebc825e194efba45008e9aea88cfb7be7db}}.",
"In the entailment context, {{cite:b7352711e0e8eb25be600fa426d221b0166821de}} introduced a large datasetwith single-evidence sentence pairs (Stanford Natural Language Inference, SNLI),but a larger vocabulary andslightly more complicated (but still conservatively formed) sentences.They also proposed baseline recurrent neural model for modelingsentence representations, while word-level attention based modelsare being studied more recently {{cite:c4632c7c532ff4e5bcd13f60de77dd1d53543aa8}} {{cite:9d0e8efc88f6450a1c85befc5c6ef696a4a9ca2b}}.",
"In the MCTest text comprehension challenge {{cite:1d878c3faa10188bd031c1449605a803e5b89b97}}, theleading models use complex engineered features ensembling multiple traditionalsemantic NLP approaches {{cite:730c4115ac16323286234b1c9011e26f247b0a08}}. The best deep model so far{{cite:b0a20425bd22db3ff2b463fce91b406890d2a28e}} uses convolutional neural networks for sentencerepresentations, and attention on multiple levels to pick evidencingsentences."
] |
[] |
Cumulative Activation in Social Networks
|
Most studies on influence maximization focus on one-shot propagation, i.e.the influence is propagated from seed users only once following a probabilisticdiffusion model and users' activation are determined via single cascade. Inreality it is often the case that a user needs to be cumulatively impacted byreceiving enough pieces of information propagated to her before she makes thefinal purchase decision. In this paper we model such cumulative activation asthe following process: first multiple pieces of information are propagatedindependently in the social network following the classical independent cascademodel, then the user will be activated (and adopt the product) if thecumulative pieces of information she received reaches her cumulative activationthreshold. Two optimization problems are investigated under this framework:seed minimization with cumulative activation (SM-CA), which asks how to selecta seed set with minimum size such that the number of cumulatively active nodesreaches a given requirement $\eta$; influence maximization with cumulativeactivation (IM-CA), which asks how to choose a seed set with fixed budget tomaximize the number of cumulatively active nodes. For SM-CA problem, we designa greedy algorithm that yields a bicriteria $O(\ln n)$-approximation when$\eta=n$, where $n$ is the number of nodes in the network. For both SM-CAproblem with $\eta<n$ and IM-CA problem, we prove strong inapproximabilityresults. Despite the hardness results, we propose two efficient heuristicalgorithms for SM-CA and IM-CA respectively based on the reverse reachable setapproach. Experimental results on different real-world social networks showthat our algorithms significantly outperform baseline algorithms.
|
[
"The classical influence maximization problem is to find a seed set of at most {{formula:07c756d3-7bd2-4ada-880c-4e54c4d38bab}} nodes to maximize the expected number of active nodes.It is first studied as an algorithmic problem by Domingos and Richardson {{cite:98e491d1d447ea27e8212a56267300866cf28e1b}}and Richardson and Domingos {{cite:d067aec4f9c920737d8fc4d8c17b2f5874bc405a}}. Kempe et al. {{cite:3fb4f93d0eb5401ed4f2932d4091ee785cae47e9}}first formulate the problem as a discrete optimization problem.They summarize the independent cascade model and the linear threshold model,and obtain approximation algorithms for influence maximization by applying submodularfunction maximization.Extensive studies follow their approach and provide more efficient algorithms{{cite:1ade17a9033db33219f5bdd57159d40e72505e35}}, {{cite:33fa132300993c55fd22b0c7e389bfad3f195d30}}, {{cite:3deb79adea4a8768b519811552f834739b57f3b2}}.Leskovec et al. {{cite:3deb79adea4a8768b519811552f834739b57f3b2}} present a “lazy-forward\" optimization method in selecting new seeds,which greatly reduce the number of influence spread evaluations. Chen et al.{{cite:1ade17a9033db33219f5bdd57159d40e72505e35}}, {{cite:33fa132300993c55fd22b0c7e389bfad3f195d30}} propose scalable algorithmswhich are faster than the greedy algorithms proposed in {{cite:8cd3cab285eeeb31b95483bcc08bfcb0192be27e}}. Recently, Borgs et al. {{cite:56734462d7f3321857f05c4370ba3e376b61c6b0}} and Tang et al. {{cite:ae49829a8c248d410356137bf989eb2ffb715bf8}}, {{cite:ef28e28ef8fabe2a9c063a5b8bca7a99ebabc0ec}} and Nguyen et al. {{cite:d411c858b1ad03b13195cb2ac4cf75e624075346}} propose a series of more effective algorithms for influence maximization in large social networks that both has theoretical guarantee and practical efficiency.The approach is based on the “Reverse Reachable Set” idea first proposed in {{cite:56734462d7f3321857f05c4370ba3e376b61c6b0}}.",
"Another aspect of influence problem is seed set minimization, Chen {{cite:038cc945c814481c26e96cecf3d19a9374917e26}} studies the seed minimizationproblem under the fix threshold model and shows some strong negative results for this model.Long et al. {{cite:e845437a915033620ae459245858dad34cefef79}} also study independent cascade model and linear threshold modelfrom a minimization perspective.In {{cite:f7c4dcf66ce3ac0eeacb03dbb2175eafbf1c50fa}}, Goyal et al. study the problem of finding theminimum size of seed set such that the expected number of active nodes reaches a given threshold,they provide a bicriteria approximation algorithm for this problem.Zhang et al. {{cite:a5adc4d4bb5aff83427891a7c1bb27aacb68717e}} study the seed set minimization problemwith probabilistic coverage guarantee, and design an approximation algorithm for this problem.He et al. {{cite:5820c254de197df20ec05728696b19d993db72e4}} study positive influence model under single-stepactivation and propose an approximation algorithm. Note that, the work in {{cite:5820c254de197df20ec05728696b19d993db72e4}} is a special case of our work.",
"Beyond influence maximization and seed minimization,another interesting direction is the learning of social influence overreal online social network data set, e.g.influence learning in blogspace {{cite:66b1b24953edf5c151d00e6d1815ab9d45673439}} and academic collaborationnetwork {{cite:bcda0373be4f7b90cd87386f9b52b99c61fbf7a3}}.",
"Most early studies on influence maximization and influence learning are summarized in the monograph {{cite:64e340dec2181d9ac49370583ad818c559359398}}.However, almost all the existing studies consider only node activation after a single information or influence cascade.Our work differentiate with all these studies on this important aspect, as discussed in the introduction.",
"Paper organization. We formally define the diffusion model and the optimization problems SM-CA and IM-CAin Section . The approximation algorithms and hardness results of these two problemsare proposed in Section , including a greedy algorithmfor SM-CA problem with {{formula:036df2b7-d429-476b-a398-f451f7c11b8b}} in section REF , the hardness result of SM-CA problem with {{formula:fa12bec5-72be-465d-9824-37c40d6537cf}} in Section REF and the inapproximate result of IM-CA problem in Section REF .In Section , we present two heuristic algorithms for SM-CA problem and two heuristic algorithms for IM-CA problem.Section shows our experimental results on real-world datasets.We summarize the paper with some further directions in Section 6."
] |
[] |
Siamese convolutional networks based on phonetic features for cognate
identification
|
In this paper, we explore the use of convolutional networks (ConvNets) forthe purpose of cognate identification. We compare our architecture with binaryclassifiers based on string similarity measures on different language families.Our experiments show that convolutional networks achieve competitive resultsacross concepts and across language families at the task of cognateidentification.
|
[
"The earlier computational effort of {{cite:e14a4d9ede864fafc2e8ae04caf33276b589535d}}, {{cite:2ce0ac0e6392fc4bda49fd35373e73edd988372d}} employsPointwise Mutual Information (PMI) to compute transition matrices between sounds.Both jager2013phylogenetic and rama2013two employ undirectional soundcorrespondence based scorer to compute word similarity. The general approach is to align word pairsusing vanilla edit distance and impose a cutoff to extract potential cognate pairs. The alignedsound symbols are then employed to compute the PMI scoring matrix that is used to realign the pairs.The PMI scoring matrix is recounted from the realigned pairs. This procedure is repeated untilconvergence.",
"jager2013phylogenetic imposes an additional cutoff based on the PMI scoring matrix.Further, jager2013phylogenetic also employs the PMI scoring matrix to infer family treesfor new language families and compares those trees with the expert trees given inGlottolog {{cite:1e27049c80287189db4778fd0880dedff7dca7ca}}. rama2013two take a slightly differentapproach, in that, the authors compute a PMI matrix independently for each language family andevaluate its performance at the task of pair-wise cognate identification. In this work, we alsocompare the convolutional networks against PMI based binary classifier.",
"Previous works of cognate identification such as {{cite:14d4116f8d35f59eea0f57c5ea0c83a8b0047f0d}}, {{cite:69da8bf5978b154cb4c63b0f59bc109ea7576d6e}} supplystring similarity measures as features for training different classifiers such as decision trees,maximum-entropy, and SVMs for the purpose of determining if a given word pair is cognate or not.",
"In another line of work, list2012sca employs a transition matrix derived fromhistorical linguistic knowledge to align and score word pairs. This approach is algorithmicallysimilar to that of Kondrak:00 who employs articulation motivated weights to score a soundtransition matrix. The weighted sound transition matrix is used to score a word pair.",
"The work of list2012sca known as Sound-Class Phonetic Alignment (SCA) approach reducesthe phonemes to historical linguistic motivated sound classes such that transitions between someclasses are less penalized than transitions between the rest of the classes. For example, theprobability of velars transitioning to palatals is a well-attested sound change across the world.The SCA approach employs a weighted directed graph to model directionality and proportionality ofsound changes between sound classes. For example, a direct change between velars and dentals isunattested and would get a zero weight. Both Kondrak:00 andlist2012sca set the weights and directions in the sound transition graph to suit thereality of sound change.",
"All the above outlined approaches employ a scoring matrix that is derived automatically or manually;or, employ a SVM to train form similarity based features for the purpose of cognate identification."
] |
[] |
Viziometrics: Analyzing Visual Information in the Scientific Literature
|
Scientific results are communicated visually in the literature throughdiagrams, visualizations, and photographs. These information-dense objects havebeen largely ignored in bibliometrics and scientometrics studies when comparedto citations and text. In this paper, we use techniques from computer visionand machine learning to classify more than 8 million figures from PubMed into 5figure types and study the resulting patterns of visual information as theyrelate to impact. We find that the distribution of figures and figure types inthe literature has remained relatively constant over time, but can vary widelyacross field and topic. Remarkably, we find a significant correlation betweenscientific impact and the use of visual information, where higher impact paperstend to include more diagrams, and to a lesser extent more plots andphotographs. To explore these results and other ways of extracting this visualinformation, we have built a visual browser to illustrate the concept andexplore design alternatives for supporting viziometric analysis and organizingvisual information. We use these results to articulate a new research agenda --viziometrics -- to study the organization and presentation of visualinformation in the scientific literature.
|
[
"Computer vision techniques have been used in the context of conventional information retrieval tasks (retrieving papers based on keyword search), including some commercial systems such as D8taplex {{cite:a91fe6da48a873b806222f5076a0fb682b9b62ae}} and Zanran {{cite:4395eecca76a1c92e5211a3ac22870a5ee54b8b3}}. Search results from these proprietary systems have not been evaluated and do not appear to make significant use of the semantics of the images.",
"In 2001, Murphy et al. proposed a Structured Literature Image Finder (SLIF) system, targeting microscope images {{cite:29e7995ff7c4633d15b480d30c03ca455bde483b}}. A decade later, Ahmed et al. {{cite:d580f33d799ed1b2c958b5a2d636daf1cddcd55a}}, {{cite:cebb6cf329214d195613904854af5c6bdcc01405}} improved the model for mining captioned figures. The latest version combines text-mining and image processing to extract structured information from biomedical literature. The algorithm first extracts images and their captions from papers, then classifies the images into six classes. Classification information and other metadata can be accessed via web service. However, SLIF focuses exclusively on microscropy images and does not extend to general figures.",
"Choudhury et al. {{cite:c56b86efc26792b3c208b1956481cec9286b428b}} proposed a modular architecture to mine and analyze data-driven visualizations that included (1) an extractor to separate figures, captions, and mentions from PDF documents {{cite:000bd161695a1be90f24c0d55d01ebf71cac0fe7}}, (2) a search engine {{cite:2d8ec8bc8aff269dc50dabd8724305ea11d05dfb}}, (3) raw-data extractor for line charts {{cite:30ddbbff16ff386902b42f5b4f988e137736f42e}}, {{cite:64f636d2331628ad187c77c6f5b492876cf0b2a9}}, {{cite:99e078c1d9917d2275020158dd4348b8f1286fd8}}, {{cite:e68927297802a6b59b8d03537a15aa72a2d63e67}}, and (4) a natural language processing module to understand the semantics of the figure. Also, they presented an integrated system from data extraction to search engine for user experience. Chen et al. {{cite:95eb6c439ce030ee7e5cef949d03deaf7eca57be}} proposed their search engine named DiagramFlyer for data-driven figures. It recovers the semantics of text components in the statistical graph. Users can search figures by giving attributes of axes or the scale range in further. Additionally, DiagramFlyer can expand queries to include related figures in terms of their production pipelines.",
"Although these early projects represent a different approach for information retrieval tasks, they make no attempt to analyze the patterns of visual information in the literature longitudinally. In this paper, we present our figure processing pipeline that classifies figure images in different categories (Section: sec:datamethod) and a search interface that uses these classified images as the primary unit of interaction to facilitate search tasks (Section: sec:ui). The key result of this paper involves the analysis of the figures in a large collection of publicly available papers (Section: sec:result).{{figure:09721157-1624-4d50-a79a-b783db0ba312}}"
] |
[] |
Re-ranking Object Proposals for Object Detection in Automatic Driving
|
Object detection often suffers from a plenty of bootless proposals, selectinghigh quality proposals remains a great challenge. In this paper, we propose asemantic, class-specific approach to re-rank object proposals, which canconsistently improve the recall performance even with less proposals. We firstextract features for each proposal including semantic segmentation, stereoinformation, contextual information, CNN-based objectness and low-level cue,and then score them using class-specific weights learnt by Structured SVM. Theadvantages of the proposed model are twofold: 1) it can be easily merged toexisting generators with few computational costs, and 2) it can achieve highrecall rate uner strict critical even using less proposals. Experimentalevaluation on the KITTI benchmark demonstrates that our approach significantlyimproves existing popular generators on recall performance. Moreover, in theexperiment conducted for object detection, even with 1,500 proposals, ourapproach can still have higher average precision (AP) than baselines with 5,000proposals.
|
[
"In recent years, object proposal has become very popular in object detection as an important pre-processing step.Object proposal methods can be classified into three main categories: window scoring based methods, grouping based methods, and CNN-based methods.",
"Window scoring based methods: Window scoring based methods attempt to score the objectness of each candidate proposal according to how likely it is to contain an object of interest.This category of methods first sample a set of candidate bounding boxes across scales and locations in an image,and measure the objectness scores based on scoring model and return top scoring candidates as proposals.Objectness {{cite:6e52f10ecbd9fa2cdc6075cc2fb830bbe7c24f5e}} is one of the earliest proposal methods.This method samples a set of proposals from salient locations in an image,and then measures objectness of each proposals according to different low-level cues, such as saliency, colour, and edges.BING {{cite:df3ea8926bbc77c82dcf552f70144be5b96b4115}} proposes a real-time proposal generator by training a simple linear SVM on binary features,the most obvious shortcoming of which is that it has a low localization accuracy.EdgeBoxes {{cite:7ba21875d4c46b9c9564e79e9a5be8b0c1c38c4d}} uses contour informations to score candidate windows without any parameter learning.In addition, it proposes a refinement process to promote localization.These methods are generally efficient, but suffer from poor localization quality.",
"Grouping based methods: Grouping based methods are segmentation-based approaches.They generally generate multiple hierarchical superpixels that are likely to contain objects for an image,and then employ different grouping strategies to generate object proposals depending on different low-level cues, such as, colour, contour and texture.Selective Search {{cite:21d682092a1d7aaddc9b9ff1292812e9b7ab52df}} greedily merges the most similar superpixels to generate proposals without learned parameters.This method has been widely applied by many state-of-the-art object detection methods.Multiscale Combinatorial Grouping (MCG) {{cite:72310c062d045533343682c3cd082d19d3bdc6e1}} generates multi-scale hierarchical segmentationsand merges them based on the edge strength to obtain object proposals.Geodesic object proposals {{cite:d450811206e7254ce98262b311ddf21e69fb2db6}} uses classifiers to place seeds for a geodesic distance transformand selects object proposal by identifying critical level sets of the distance transforms.Compared to window scoring based methods,grouping based methods have better localization ability, but require more complex computation.",
"CNN-based methods: Benefit from the strong discrimination ability of Convolutional Neural Network (CNN), CNN-based methods directly generate high quality candidate proposals with a fully convolutional network (FCN).Multibox {{cite:b201e98ed6e9be186e32a2cfe2667acb20b2a1eb}} trains a large CNN model to directly generate object proposals from images and ranks them depending on the predicted objectness scores.RPN {{cite:fed131141b238b0277e5f73a46e417f8fbda80d4}} uses a FCN to generate object proposals with a wide range of scales and aspect ratios.DeepBox {{cite:cd8cd13403abbfd01d5b6c72718fd925432785d9}} uses a lightweight CNN to predict the objectness scores of candidate proposals and re-ranks them depending on the predicted objectness scores.These CNN-based methods achieve high recall with only a small number of proposals under loose criteria (e.g. IoU > 0.5),but fails under strict criteria (e.g. IoU > 0.7)."
] |
[] |
Automatic Construction of Discourse Corpora for Dialogue Translation
|
In this paper, a novel approach is proposed to automatically constructparallel discourse corpus for dialogue machine translation. Firstly, theparallel subtitle data and its corresponding monolingual movie script data arecrawled and collected from Internet. Then tags such as speaker and discourseboundary from the script data are projected to its subtitle data via aninformation retrieval approach in order to map monolingual discourse tobilingual texts. We not only evaluate the mapping results, but also integratespeaker information into the translation. Experiments show our proposed methodcan achieve 81.79% and 98.64% accuracy on speaker and dialogue boundaryannotation, and speaker-based language model adaptation can obtain around 0.5BLEU points improvement in translation qualities. Finally, we publicly releasearound 100K parallel discourse data with manual speaker and dialogue boundaryannotation.
|
[
"In the specific case of dialogue MT system, data acquisition can impose challenges including data scarcity, translation quality and scalability. The release of the Penn Discourse Treebank (PDTB)Available at https://www.seas.upenn.edu/~pdtb. {{cite:991f08ce73e0b4fc9f5acb44b6f279be25b9bb5f}} helped bring about a new sense of maturity in discourse analysis, finally providing a high-quality large-scale resource for training discourse parsers for English. Based on PDTB, some have applied the insights to MT {{cite:480d25eb73ada9692e201f83491cf9bcdbc28fea}}. A resource like the PDTB is extremely valuable, and it would be desirable to have a similar resource in dialogue or conversation as well.",
"There are two directions of work related to dialogue corpus construction. One is parallel corpora construction for dialogue or conversation MT {{cite:3d8cd6197da17c4ddf60f31634ffd0cd503449b6}}, {{cite:b93526019d8274c1622a0025aa5053d7ce4f70cb}}, {{cite:aba375a5965f4ad02581ba79cda2d64a0846a317}}, {{cite:d9c5ee68680f4590db3fad7d72bca0cadc8a549e}}, {{cite:669539fa717a196eef9870097376b464174c91f2}}, {{cite:5f71e833bec82965291c9a07bc851b7e63683649}}, {{cite:895dd2dd8451b76306bfc918cc9056356ee833a4}}. Thanks to the effects of crowdsourcing and fan translation in audiovisual translation {{cite:157ae04bd39a137124d43f503ec59d14aed8784a}}, we can regard subtitles as parallel corpora. zhang2014dual leveraged the existence of bilingual subtitles as a source of parallel data for the Chinese-English language pair to improve the MT systems in the movie domain. However, their work only considers sentence-level data instead of extracting more useful information for dialogues. Besides, Japanese researchers constructed a speech dialogue corpus for a machine interpretation system {{cite:06c916954e8b9e4cf905c06fa2653b5a3c6de774}}, {{cite:1bd13b836228afc119c520a06c2fd2347b548c4c}}, {{cite:59a3f7040026fcd6db52ca5cb3a9eb5be99d1960}}, {{cite:f8046578b7d8be2dab4edb01de4fe317822dfb4f}}. They collected speech dialogue corpora for machine interpretation research via recording and transcribing Japanese/English interpreters' consecutive/simultaneous interpreting in the booth. The German VERBMOBIL speech-to-speech translation programme {{cite:8ce4c64c158dcc7520b8a4895e665ecce537582c}} also collected and transcribed task-oriented dialogue data. This related work focused on speech-to-speech translation including three modules of automatic speech recognition (ASR), MT and text-to-speech(TTS).",
"The other one is mining rich information from other resources such as movie scripts. danescu2011chameleons created a conversation corpus containing large metadata-rich collections of fictional conversations extracted from raw movie scripts. Both banchs2012movie and CMU released dialogue corpora extracted from the Internet Movie Script Database (IMSDb).Available at http://www.imsdb.com.Based on IMSDb, walker2012annotated annotated 862 film scripts to learn and characterize the character style for an interactive story system, and schmitt2012parameterized annotated 347 dialogues to explore a spoken dialogue system. The resource of movie scripts, such as IMSDb, is good enough to generate conversational discourse for dialogue processing. However, monolingual movie scripts are not enough for MT which needs a large-scale bilingual dialogue corpus to train and tune translation models."
] |
[] |
Mixtape Application: Last.fm Data Characterization
|
This report analyses data collected from Last.fm and used to create areal-time recommendation system. We collected over 2M songs and 1M tags and372K user's listening habits. We characterize users' profiles: age, playcount,friends, gender and country. We characterized song, artist and tag popularity,genres of songs. Additionally we evaluated the co-occurrence of songs in users'histories, which can be used to compute similarity between songs.
|
[
"The most relevant aspects of recommendation systems are the premises theyare based on, and how they build upon them. One of the most common premisesis content similarity, and that it would be desirable to recommend contentthat is somehow similar to content that a given user enjoys. So, We presenta small compilation of related works in the area of Media Similarity,Music Recommendation and Playlist Generation.",
"Media item-to-item similarity computation:Several research studies offer technical solutions to define similarity between a given pair of items, and they divide themselves among two schools. Some studies develop similarity measures that are based solely on content (objective approach), being that audio or video information — for instance, the spectral or rhythmic content of songs.On the other hand, other studies develop similarity measures that are based on user-generateddata and tags, also known as collaborative filtering (subjective approach). In the context of music,different approaches to define item-to-item similarity have been studied extensively, such ascontent-based measures {{cite:23cb1bc901f7b028b1658093342a8db326fdc554}}, {{cite:f7e4abd044e570b2da51dbed040471c3848d5d95}}, {{cite:e9327345c9a09d201aa25da48ce4c5ff7eda70fd}}, {{cite:b4b225074727a195f42033dd26d8f935307db41c}}, {{cite:b6d1f61c9bafa52340c6900318712451520d7d56}}, {{cite:ebdd90e81baa962a0c3b546af1566f368ae096f9}},that analyze spectral or rhythmic properties of songs, and user-basedmeasures {{cite:75e2023e967ac111b7441fbf4dddf62691275f34}}, {{cite:a1665f48ec08287dc4378aab6139e39728331b97}}, {{cite:27344a3266e1c215b13d5ba1809e7d0e4bf3d1e0}}, {{cite:27a824eafb460ae85c5099c7030845a0530a46b2}}, {{cite:17d942ff0e66beb91da75d4290fbb07d161103c2}}, {{cite:df3e1bb2db5bcac317323148c4c328ef4a3bb2c7}}, thatanalyze user listening habits in online social networks or user-generatedtags {{cite:27344a3266e1c215b13d5ba1809e7d0e4bf3d1e0}}.Many of these studies aim at building recommendationsystems {{cite:23cb1bc901f7b028b1658093342a8db326fdc554}}, {{cite:f7e4abd044e570b2da51dbed040471c3848d5d95}}, {{cite:ebdd90e81baa962a0c3b546af1566f368ae096f9}}, {{cite:17d942ff0e66beb91da75d4290fbb07d161103c2}}, {{cite:4b4af85b335a89c274186dbbd7949cc16557114a}}, {{cite:27344a3266e1c215b13d5ba1809e7d0e4bf3d1e0}}.",
"Music playlist generation: Another related line of research to musicnavigation systems is playlist generation. Several people haveaddressed this problem from different perspectives. There are techniques thatuse statistical analysis of radio streams {{cite:42a9c1f6c83c775424a71b7a94003fb16e865a04}}, {{cite:4b4af85b335a89c274186dbbd7949cc16557114a}}, {{cite:70610b305326e51f04d991156a9bcdb6d1e89820}}, {{cite:1eb5d328ed2fc5a7fd80847cfc880b4b3c84f711}}, are based onmultidimensional metric spaces {{cite:1eb5d328ed2fc5a7fd80847cfc880b4b3c84f711}}, {{cite:27a824eafb460ae85c5099c7030845a0530a46b2}}, {{cite:e09c73777fbf972a828c7925c327cdfb3d0955a1}}, {{cite:57acd2ec2fe17f833bf7482dc4ce2c1d7062640a}}, {{cite:df3e1bb2db5bcac317323148c4c328ef4a3bb2c7}}, explore audiocontent {{cite:787179ce68f5f72919dba3fa2d74b1905835a8cf}},and user skipping behavior {{cite:556e5936706dcd2ffdb414c9a8fb32b0f3c2c67c}}.",
"In particular, authors in {{cite:556e5936706dcd2ffdb414c9a8fb32b0f3c2c67c}} create playlistsbased on audio music similarity and skipping behavior, while authors in{{cite:3f2d7f613282b30f623671a0e1b52c9fab644a31}} use network flow analysis to generate playlists from afriendship graph of artists on MySpace. Maillet et al {{cite:42a9c1f6c83c775424a71b7a94003fb16e865a04}},in turn, present an approach to generating steerable playlists from tags linkedto songs played in professional radio station playlists, and Chen et al {{cite:1eb5d328ed2fc5a7fd80847cfc880b4b3c84f711}}model playlists as Markov chains, which are generated through the Latent MarkovEmbedding machine learning algorithm. The heuristics in {{cite:556e5936706dcd2ffdb414c9a8fb32b0f3c2c67c}}use acoustic similarity and present linear complexity to return the nextitem to the user, which is feasible to navigate in a collection of 2,500 items,as the authors did, but does not scale to larger collections, which is not onlydesirable but necessary for modern standards."
] |
[] |
Low-Cost Scene Modeling using a Density Function Improves Segmentation
Performance
|
We propose a low cost and effective way to combine a free simulation softwareand free CAD models for modeling human-object interaction in order to improvehuman & object segmentation. It is intended for research scenarios related tosafe human-robot collaboration (SHRC) and interaction (SHRI) in the industrialdomain. The task of human and object modeling has been used for detectingactivity, and for inferring and predicting actions, different from those works,we do human and object modeling in order to learn interactions in RGB-D datafor improving segmentation. For this purpose, we define a novel densityfunction to model a three dimensional (3D) scene in a virtual environment(VREP). This density function takes into account various possibleconfigurations of human-object and object-object relationships and interactionsgoverned by their affordances. Using this function, we synthesize a large,realistic and highly varied synthetic RGB-D dataset that we use for training.We train a random forest classifier, and the pixelwise predictions obtained isintegrated as a unary term in a pairwise conditional random fields (CRF). Ourevaluation shows that modeling these interactions improves segmentationperformance by ~7\% in mean average precision and recall over state-of-the-artmethods that ignore these interactions in real-world data. Our approach iscomputationally efficient, robust and can run real-time on consumer hardware.
|
[
"For vision applications, we can acquire RGB-D data using real-world sensors or by generating synthetic data using computer graphics rendering techniques in a 3D environment. The real-world data is also possible to use for training the system, but researchers generally avoid this because of several reasons: real-world data generation usually requires expensive equipments (i.e. software and hardware gadgets). In addition manual hand labeling of objects is a very-time consuming and tedious task because precise measurement of objects is required. Also manual ground truth labeling leads to missing out several important information in the scene due to imperfect and error-prone human annotations. Therefore, some vision researchers avoid using real-world data for training vision systems. While by using the present precise rendering techniques and simulation frameworks with modeled environment, it is possible to synthesize accurate and realistic data, which is very close to real-world data. This provides the flexibility to tune the elements of scene for incorporating camera noise, lighting-effects, shadows, rotation/translation, etc. during rendering in order to make the synthetic data as consistent as possible with the physical real-world data.",
"Here, we discuss some of the previous works related to vision problems. In {{cite:4e3f651cd7abe95c46e900826e347b34ccd82e7a}}, Neumann-Cosel et al. use synthetic data obtained from the modeled Virtual Test Drive simulation for lane tracking in driver assistance and active safety systems. An identical work from Haltakov et al. {{cite:3cf877a5fbdfde502b07e7a0c5bf27dff8d86f93}} proposes a framework for generation of synthetic data from a realistic driving simulator to create modeled traffic scenarios similar to real-world scenarios. A number of works uses synthetic human data for pose and activity recognition {{cite:01df8f73ce04c3fd0de387bb4d184e06c2bddcd4}}, {{cite:944968616513bd44b172a22f85d9e946aec8f231}}, {{cite:b3d952851af0b0334f34111b6c3346a68054cd60}}, {{cite:ac4026f2c2d35890652033c23354b425e6be6b7e}}, {{cite:752434c0479bda776f14f736a5df587c8294fd30}}, {{cite:fe178f4d110af7de65b2c72e2f069752066a6021}}, {{cite:bc1c48a5ad90a17ff0e4d59d94c4d447c6001813}}. In vision problems, human pose and activity recognition has received great attention in the last decade. In {{cite:fe178f4d110af7de65b2c72e2f069752066a6021}}, Shotton et al. synthesize a large and varied dataset of human actions using a motion capture technique. In {{cite:944968616513bd44b172a22f85d9e946aec8f231}}, {{cite:752434c0479bda776f14f736a5df587c8294fd30}}, the authors use a simple synthetic human body representation in a virtual environment using a KINECT skeleton estimation for generating synthetic human pose data. In {{cite:fe178f4d110af7de65b2c72e2f069752066a6021}}, the data is generated based on front, top, and side view, while in {{cite:944968616513bd44b172a22f85d9e946aec8f231}}, {{cite:752434c0479bda776f14f736a5df587c8294fd30}}, it is top view only. Both works are applied in real-world scenarios, but neither of the models incorporate the physical interaction of objects with other objects. Also no structural and modeled arrangements are dealt with. Thus, their algorithm fails to recognize human body-parts, when human body-parts overlap other objects, or are partially occluded by other objects in the scene. In {{cite:1541882bdbb816fdd8268767f86689ac7baada92}}, Hoiem et al. discard the object candidates that are not well supported by proper occlusion boundaries. Yet, we know that reasoning about occlusion would definitely improve the recognition performance. A substantial little work has been done in the context of scene modeling.",
"Our approach is driven by three key objectives namely computational efficiency, robustness and time efficiency (i.e. real-time) for industrial applications. It further differs from Shotton et al. {{cite:fe178f4d110af7de65b2c72e2f069752066a6021}} in the following aspects. (a) In {{cite:fe178f4d110af7de65b2c72e2f069752066a6021}}, all training data were thereby synthetically generated by using marker based motion capture (Mo-Cap) technique. Such setups are very expensive and inaccessible to most users. On the other hand, we use a highly optimized virtual representation of 3D human skeleton in a virtual environment. We generate the synthetic data of human body-parts in a virtual environment using a multi-sensor KINECT setup for skeleton tracking in real-world {{cite:944968616513bd44b172a22f85d9e946aec8f231}}. In addition to human data, we also generate synthetic data for objects as well. (b) Calibration of Mo-Cap is a tedious task in comparison to KINECT setup. (c) Processing Mo-Cap data requires sophisticated hardware in comparison to KINECT depth data which can be porcessed with consumer hardware. This way computational expense is reduced.",
"For learning and reasoning about the physical scene, we need to capture the relevant and meaningful information (i.e. geometry, pose, shape, occlusion, and other attributes) and train the classifier system. The density functions capture the relationship, physical interactions, and the context among objects as well as the geometry of the objects. For physical scene understanding, occlusion boundaries should be well identified, since they form a very crucial component for depth ordering. Incorporating occlusion during synthetic scene modeling, helps to get a better understanding of the real-world scene. In {{cite:ac4026f2c2d35890652033c23354b425e6be6b7e}}, Jiang et al. model a density function based on 3D spatial features to capture the object-object and human-object relationship in a 3D scene. They use the density function to infer arrangements of object in a scene. In {{cite:ac4026f2c2d35890652033c23354b425e6be6b7e}}, a camera is installed in the virtual environment, in the corner of the room, where two walls and the ceiling meet, or where a wall and the ceiling meet. Their evaluation is only limited to the virtual environment for training and testing assessments. Similar is done in {{cite:01df8f73ce04c3fd0de387bb4d184e06c2bddcd4}}, where Dittrich et al. model the density function on depth features using a top-view RGB-D sensor. Their evaluation is also limited to their virtual environment for assessment. Our work is inspired by top-view segmentation from Shotton et al. {{cite:fe178f4d110af7de65b2c72e2f069752066a6021}} and Sharma et al. {{cite:752434c0479bda776f14f736a5df587c8294fd30}}, and we define the density function in VREP {{cite:95fb2b4dfedea19306eb1be3794e5cec38da775b}} in order to model object-object and human-object relationships. We use a top-view RGB-D sensor data. However, our evaluation is not just limited to virtual environment, but also to physical real-world scenes. In addition, our work focuses on modeling relationships between a group of multiple 3D objects in a scene, while Shotton et al. {{cite:fe178f4d110af7de65b2c72e2f069752066a6021}} and Sharma et al. {{cite:752434c0479bda776f14f736a5df587c8294fd30}} only focuses on a single 3D object in a scene. Moreover, they do not incorporate any form of interaction modeling (i.e. non-modeled) between humans and objects.",
"Some other work, similar to our work, Gupta et al. {{cite:9afab0ded23e4e8e63ab4892045580657053847f}}, Aksoy et al. {{cite:842176a09462541b3c5b5823ac62f9c564d22dd9}}, and Pieropan et al. {{cite:0fb09d34b83cdac80f2f58b39e37b724fccc3f1f}} use the spatial relationship to perform activities recognition where human-object and object-object relationships are encoded. Also some other notable work, similar to ours that uses the same idea of output of a classifier as input of another model are like “Autocontext\" by Tu et al. {{cite:a0a00e3ba24993d519d18ea9835cc6c47a19f3ef}}, “Decision Tree Fields\" by Nowotzin et al. {{cite:de789e520de75123fb445668bcea53d8d4e5adb9}} or “Structured Class-Labels in Random Forests for Semantic Image Labelling\" by Kontschieder et al. {{cite:e5672c1c7bdbd1f3a3d6afded66a21185a58d08e}}, among others.",
"One major problem that arises when using synthetic data, is how to determine if the synthesized data is realistic in comparison to real-world data. This could be determined by evaluating the statistics of object constellation in the synthetic and real world. A simple approach could be to apply collision avoidance while modeling interactions between objects, but such a modeling is not realistic as occlusions never exist in the modeled scene. Therefore, a realistic scene with appropriate object interactions could be modeled using a density function in order to produce good prediction results, not just for synthetic data, but also for real-world data (see Section and Section REF for details)."
] |
[] |
Estimation of Passenger Route Choice Pattern Using Smart Card Data for
Complex Metro Systems
|
Nowadays, metro systems play an important role in meeting the urbantransportation demand in large cities. The understanding of passenger routechoice is critical for public transit management. The wide deployment ofAutomated Fare Collection(AFC) systems opens up a new opportunity. However,only each trip's tap-in and tap-out timestamp and stations can be directlyobtained from AFC system records; the train and route chosen by a passenger areunknown, which are necessary to solve our problem. While existing methods workwell in some specific situations, they don't work for complicated situations.In this paper, we propose a solution that needs no additional equipment orhuman involvement than the AFC systems. We develop a probabilistic model thatcan estimate from empirical analysis how the passenger flows are dispatched todifferent routes and trains. We validate our approach using a large scale dataset collected from the Shenzhen metro system. The measured results provide uswith useful inputs when building the passenger path choice model.
|
[
"Building users route choice model is an important research direction in the field of transportation {{cite:34cc528180047c8c7453330e7c196740c567a760}}, which is the basis for traffic management policies-making. Due to the lack of the observation of how probably each route is chosen for an OD pair with multiple routes, most of the past studies focus on building route choice models from empirical perspective. They assume that all passengers have full knowledge of the transportation when attempting to minimize some objective functions e.g., minimizing their travel time (user equilibrium) or minimizing the total system travel time (system optimum){{cite:1dc1460850b2ec4f87a0c50762e3e9f77f9375a5}}, {{cite:5320175ee8c8a7e6086dd26d1b75dcdd19f56296}}, {{cite:0c93cd80373a7693b31e5f634b0775f967982f07}}. However, those models depend heavily on behavior assumptions and lack in reliable supporting data. Given the dynamic and stochastic nature of transportation systems, the assumption of the passengers' global knowledge is questionable.",
"Fortunately, the large amount of smart card data in a long period provide us a great opportunity to analyze passengers' transit behavior and evaluate transit service. There were a few previous studies regarding the utilization of smart card data. The literature {{cite:30367b8624a0415f0eaf88ffa66b5d9aba7792d2}} considered the potential usage of smart card data for travel. The literature {{cite:59293e285e4f8d2c90be654bf24770f01d32f2ce}} analyzed users' travel behavior using data mining technology, which clustered users into four groups according to their temporal travel patterns. Our recent work {{cite:fbfa339ea09e27044c61d2bdfb9c251e66c58640}} studied individual passenger's temporal and spatial travel patterns. We found that if a passenger is temporally regular, it is very possible that the passenger is also spatially regular. Besides understanding travel behavior, smart card data have been used to improve public transit services. The study in {{cite:45db53a6ae5d689c43db025178ddec22a3b65d22}} gave a comprehensive review of using smart card data from different aspects: strategic, tactical and operational. To improve the resilience to service disruptions of metro systems, paper {{cite:5de73af0fff82ac6a080ee28a2e4995ca3f3f067}} investigated a practical problem about integrating localized bus service with metro network. Using the same data set, three optimization models were formulated to design demand-driven timetables for a single-track metro service {{cite:cc49b4cf91da3b6c6f4dea9cac41083eded3c900}}.",
"For understanding passengers' flow assignments in metro system, The authors in {{cite:7e0b3a3ba53d78b20d3a5db5058e1afa22b6a8b3}} proposed a method to estimate which trains every smart card holder boarded during his/her journey. This method could be used to estimate trains' occupancies. However, it was also based on some assumptions that may be only available in some limited scenarios: (a)The methodology assumed that all passengers know the train timetable beforehand. When choosing route, they will first choose the plan with minimum total waiting time, then choose the plan with fewer transits. The remaining small percentage of undecided trips were assumed to be assigned to all possible plans with an equal probability. (b)The walking time between the charge gates and that platform, and the walking time for transfer between platforms are ignored, which may lead to mismatching between passengers' tap time and trains' operation time.",
"The authors in {{cite:c0bdba2962e2762f8c39ea43b5df3ee4f761bfac}} proposed a linear regression model to analyze the individual trajectory during a metro trip, which could be used to estimate the spatial-temporal density inside a metro system. However, their model focused on a single-track scenario that is oversimplified. The study in {{cite:3585cb0838bfa7576defccfd03e2e63bb5e736c1}} used a clustering algorithm to infer the route-use patterns of metro passengers from the smart-card data. It confirmed that a Gaussian mixture model worked well in finding the route shares and the mean and variance of travel times for each route of London underground. But the conclusion based on two preconditions. First, the number of routes used by users must be known in advance. Second, the probability distribution function of travel time of each route must be Gaussian. The study in {{cite:9564209c71cf852470ce833c58d6bf3d100c2157}} developed an integrated Bayesian approach to infer both network attributes and passenger route choice behavior in a complex metro system, which worked well in some cases that there are lack of train timetable. But a large set of explanatory variables and the probability distribution of these variables need to be calibrated, such as it assumed that all link costs are characterized by independent normal distributions. This is not always true. Taking the phenomenon “left-behind\" as example, the imbalance of geographical distribution of passengers leads to the various “left-behinds” in different stations and results that the time cost does not follow normal distributions in a station for different number of trains that need to be waited for. There are other related work of big data based analysis for smart transportation systems {{cite:1f873a722b685106c8991ce5f9e3dd8f86e7b487}}, {{cite:b6cbb164147ec810c101a2e0e4b1d98b295c7f22}}, {{cite:f27a488bc6df571e4aa659a9c5fc354ba11a0a19}}, {{cite:f353cd967b18643233e1b9849319f8a352ff8d1c}}, while they are not targeting metro systems.",
"In sum, the existing methods did not consider the passengers' “left-behind” in detail, which however is one of the main factors affecting us to understanding passengers' path choice behavior. In this paper, we propose a novel approach to calculate the probability of each route used for an OD with multiple routes, it can be used to complicated traffic situation of complex metro network, especially for the situation that “left-behind” occurs often."
] |
[] |
ObliviSync: Practical Oblivious File Backup and Synchronization
|
Oblivious RAM (ORAM) protocols are powerful techniques that hide a client'sdata as well as access patterns from untrusted service providers. We present anoblivious cloud storage system, ObliviSync, that specifically targets one ofthe most widely-used personal cloud storage paradigms: synchronization andbackup services, popular examples of which are Dropbox, iCloud Drive, andGoogle Drive. This setting provides a unique opportunity because the aboveprivacy properties can be achieved with a simpler form of ORAM calledwrite-only ORAM, which allows for dramatically increased efficiency compared torelated work. Our solution is asymptotically optimal and practically efficient,with a small constant overhead of approximately 4x compared with non-privatefile storage, depending only on the total data size and parameters chosenaccording to the usage rate, and not on the number or size of individual files.Our construction also offers protection against timing-channel attacks, whichhas not been previously considered in ORAM protocols. We built and evaluated afull implementation of ObliviSync that supports multiple simultaneous read-onlyclients and a single concurrent read/write client whose edits automatically andseamlessly propagate to the readers. We show that our system functions underhigh work loads, with realistic file size distributions, and with smalladditional latency (as compared to a baseline encrypted file system) whenpaired with Dropbox as the synchronization service.
|
[
"ORAM. ORAM protects the access pattern from an observer such that itis impossible to determine which operation is occurring, and on whichitem. The seminal work on the topic is by Goldreich andOstrovsky {{cite:95b3a403f3969a93e73032961e7fa2ad2a699b57}}, and since then, many works have focused onimproving efficiency of ORAM in both the space, time, and communicationcost complexities (for example {{cite:d7131733565c0d4989652bfeb9553d671d20b3e8}}, {{cite:23eb96df89562b4db5e7f2df4decc7bbac2df55d}}, {{cite:1baeb99367eb26609a9af1d1cfe8245fa30ea3ab}}, {{cite:c7c157f4cd34014b3b4c2e4d3a51c94df7544594}}, {{cite:ed7245e606e2b78bcffcfca81e7ee60f6c1e999e}}just to name a few; see the references therein).Blass et al. introduced write-only ORAMs {{cite:2850881f1dc64a9cce79cded15ab23c77742ba52}}. In awrite-only ORAM, any two write accesses are indistinguishable, and theyachieved a write-only ORAM with optimal {{formula:c8615ffa-c1b6-4b77-962b-7874150cd772}} communication complexityand only poly-logarithmic user memory.Based on their work, we construct a write-only ORAM that additionallysupports variable-size data and hides the when the data items aremodified. We point out also that variable-sized blocks in traditionalread/write ORAMs were also considered recently by {{cite:18266a66061cc67caf6675441db871b1a5092350}}, but withhigher overhead than what can be achieved in the write-only setting.",
"Protecting against timing side-channels. Side-channel attacks that use network traffic analysis in order to learnprivate information have been considered in contexts other than securecloud storage. Proposed systems for location tracking {{cite:032113d3713ae8eb4b1229f4a89b31ea368269dd}} andsystem logging {{cite:bc685a341d75d10172f6f07b73b2e9c2c768f3c0}} use buffering and random or structureddelays to protect against such attacks in a similar way to our work.",
"Personal cloud storage. A personal cloud storage offers automatic backup, file synchronization,sharing and remote accessibility across a multitude of devices andoperating systems.Among the popular personal cloud storages are Dropbox, Google Drive, Box,and One Drive.However, privacy of cloud data is a growing concern, and to address thisissue, many personal cloud services with better privacy appeared. Amongthe notable services are SpiderOak {{cite:3c3a34f68382bab829ad57af96cee810f1a62dac}},Tresorit {{cite:a1db39fd612e65755a2fc920d717586e1ed7772b}}, Viivo {{cite:5048384995804cc7d15b5d27e8952b4b657ef3bd}},BoxCryptor {{cite:c8a150755e22fb8af49adf9eb96586a4b68cf380}}, Sookas {{cite:383d3d3a89f5cd3868612d5768f4f3fd28fa2a65}},PanBox {{cite:e3b8caa7904940533c581a67f6edaa9bc3e922dc}}, and OmniShare {{cite:245af4c64884207a7a9a67d6e938a3efcccac10b}}.All the solutions achieve better privacy by encrypting the file datausing encryption keys created by the client.We stress that however there has been no attempt to achieve the strongerprivacy guarantee of obliviousness."
] |
[] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.