Datasets:

---

allenai

/

scitldr

like 28

Follow

Ai2 2.98k

[ "We propose a novel unsupervised generative model, Elastic-InfoGAN, that learns to disentangle object identity from other low-level aspects in class-imbalanced datasets.", "We first investigate the issues surrounding the assumptions about uniformity made by InfoGAN, and demonstrate its ineffectiveness to properly disentangle object identity in imbalanced data.", "Our key idea is to make the discovery of the discrete latent factor of variation invariant to identity-preserving transformations in real images, and use that as the signal to learn the latent distribution's parameters.", "Experiments on both artificial (MNIST) and real-world (YouTube-Faces) datasets demonstrate the effectiveness of our approach in imbalanced data by:", "(i) better disentanglement of object identity as a latent factor of variation; and", "(ii) better approximation of class imbalance in the data, as reflected in the learned parameters of the latent distribution.", "Generative models aim to model the true data distribution, so that fake samples that seemingly belong to the modeled distribution can be generated (Ackley et al. (1985) ; Rabiner (1989) ; Blei et al. (2003) ).", "Recent deep neural network based models such as Generative Adversarial Networks (Goodfellow et al. (2014) ; Salimans et al. (2016) ; ) and Variational Autoencoders (Kingma & Welling (2014) ; Higgins et al. (2017) ) have led to promising results in generating realistic samples for high-dimensional and complex data such as images.", "More advanced models show how to discover disentangled representations ; Chen et al. (2016) ; Tran et al. (2017) ; Hu et al. (2018) ; Singh et al. (2019) ), in which different latent dimensions can be made to represent independent factors of variation (e.g., pose, identity) in the data (e.g., human faces).", "InfoGAN ) in particular, tries to learn an unsupervised disentangled representation by maximizing the mutual information between the discrete or continuous latent variables and the corresponding generated samples.", "For discrete latent factors (e.g., digit identities), it assumes that they are uniformly distributed in the data, and approximates them accordingly using a fixed uniform categorical distribution.", "Although this assumption holds true for many existing benchmark datasets (e.g., MNIST LeCun (1998)), real-word data often follows a long-tailed distribution and rarely exhibits perfect balance between the categories.", "Indeed, applying InfoGAN on imbalanced data can result in incoherent groupings, since it is forced to discover potentially non-existent factors that are uniformly distributed in the data; see Fig. 1 .", "In this work, we augment InfoGAN to discover disentangled categorical representations from imbalanced data.", "Our model, Elastic-InfoGAN, makes two modifications to InfoGAN which are simple and intuitive.", "First, we remodel the way the latent distribution is used to fetch the latent variables; we lift the assumption of any knowledge about class imbalance, where instead of deciding and fixing them beforehand, we treat the class probabilities as learnable parameters of the optimization process.", "To enable the flow of gradients back to the class probabilities, we employ the Gumbel-Softmax distribution (Jang et al. (2017) ; Maddison et al. (2017) ), which acts as a proxy for the categorical distribution, generating differentiable samples having properties similar to that of categorical samples.", "Second, we enforce our network to assign the same latent category for an image I and its transformed image I , which induces the discovered latent factors to be invariant to identity-preserving transformations like illumination, translation, rotation, and scale changes.", "Although there are multiple meaningful ways to partition unlabeled data-e.g., with digits, one partitioning could be based Samples generated with an InfoGAN model learned with a fixed uniform categorical distribution Cat(K = 10, p = 0.1) on balanced and imbalanced data, respectively.", "Each row corresponds to a different learned latent category.", "(Right): Samples generated with Elastic-InfoGAN using its automatically learned latent categorical distribution.", "Although InfoGAN discovers digit identities in the balanced data, it produces redundant/incoherent groupings in the imbalanced data.", "In contrast, our model is able to discover digit identities in the imbalanced data.", "on identity, whereas another could be based on stroke width-we aim to discover the partitioning that groups objects according to a high-level factor like identity while being invariant to low-level \"nuisance\" factors like lighting, pose, and scale changes.", "Such partitionings focusing on object identity are more likely to be useful for downstream visual recognition applications (e.g., semi-supervised object recognition).", "In sum, our modifications to InfoGAN lead to better disentanglement and categorical grouping of the data (Fig. 1) , while at the same time enabling the discovery of the original imbalance through the learned probability parameters of the Gumbel softmax distribution.", "Importantly, these modifications do not impede InfoGAN's ability to jointly model both continuous and discrete factors in either balanced or imbalanced data scenarios.", "Our contributions can be summarized as follows: (1) To our knowledge, our work is the first to tackle the problem of unsupervised generative modeling of categorical disentangled representations in imbalanced data.", "We show qualitatively and quantitatively our superiority in comparison to Info-GAN and other relevant baselines.", "(2) Our work takes a step forward in the direction of modeling real data distributions, by not only explaining what modes of a factor of variation are present in the data, but also discovering their respective proportions.", "In this work, we proposed a new unsupervised generative model that learns categorical disentanglement in imbalanced data.", "Our model learns the class distribution of the imbalanced data and enforces invariance to be learned in the discrete latent variables.", "Our results demonstrate superior performance over alternative baselines.", "We hope this work will motivate other researchers to pursue this interesting research direction in generative modeling of imbalanced data." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.15789473056793213, 0.1463414579629898, 0.1818181723356247, 0.2222222238779068, 0.13793103396892548, 0.1875, 0.08510638028383255, 0.06896550953388214, 0.1666666567325592, 0.1395348757505417, 0.1304347813129425, 0.0833333283662796, 0.260869562625885, 0.32258063554763794, 0.06666666269302368, 0.15686273574829102, 0.15094339847564697, 0, 0.10169491171836853, 0.07692307233810425, 0.06896550953388214, 0.25, 0.25806450843811035, 0.07843136787414551, 0, 0.11999999731779099, 0.14999999105930328, 0.31111109256744385, 0.06451612710952759, 0.16326530277729034, 0.29411762952804565, 0.277777761220932, 0, 0.2222222238779068 ]

[ "Elastic-InfoGAN is a modification of InfoGAN that learns, without any supervision, disentangled representations in class imbalanced data" ]

[ "Many real applications show a great deal of interest in learning multiple tasks from different data sources/modalities with unbalanced samples and dimensions.", "Unfortunately, existing cutting-edge deep multi-task learning (MTL) approaches cannot be directly applied to these settings, due to either heterogeneous input dimensions or the heterogeneity in the optimal network architectures of different tasks.", "It is thus demanding to develop knowledge-sharing mechanism to handle the intrinsic discrepancies among network architectures across tasks.", "To this end, we propose a flexible knowledge-sharing framework for jointly learning multiple tasks from distinct data sources/modalities.", "The proposed framework allows each task to own its task (data)-specific network design, via utilizing a compact tensor representation, while the sharing is achieved through the partially shared latent cores.", "By providing more elaborate sharing control with latent cores, our framework is effective in transferring task-invariant knowledge, yet also being efficient in learning task-specific features.", "Experiments on both single and multiple data sources/modalities settings display the promising results of the proposed method, especially favourable in insufficient data scenarios.", "Multi-task learning (MTL) (Caruana, 1997; Maurer et al., 2016) is an approach for boosting the overall performance of each individual task by learning multiple related tasks simultaneously.", "In the deep learning setting, jointly fitting sufficiently flexible deep neural networks (DNNs) to data of multiple tasks can be seen as adding an inductive bias to the deep models, which can facilitate the learning of feature representations that are preferable by all tasks.", "Recently, the deep MTL has been successfully explored in a broad range of applications, such as computer vision (Zhang et al., 2014; Misra et al., 2016) , natural language processing (Luong et al., 2015; , speech recognition Huang et al., 2015) and so on.", "Nevertheless, one key challenge in deep MTL remains largely unaddressed, that is, almost all existing deep MTL approaches (Yang & Hospedales, 2017; Long et al., 2017) restrict themselves only to the setting of multi-label learning (or multi-output regression) (Zhang & Yang, 2017) .", "In other words, different tasks must be fed with input data from the same source (or domain).", "Such requirement, however, seriously limits the applicability of those models to a more realistic scenario of deep MTL, where the tasks involve distinct data sources (domains) with unbalanced sample sizes or dimensions.", "More specifically, tasks from some domains with abundant samples or small input dimensions are relatively easy to handle, whereas tasks from other domains are quite challenging due to the insufficient training data and large dimensionality.", "For instance, classifying hand-written digits (MNIST dataset (LeCun et al., 1998) ) is somewhat similar to the recognition of hand-drawn characters (Omniglot dataset (Lake et al., 2015) ).", "The Omniglot task is much harder than the MNIST task, as each character in Omniglot has only 20 training samples, while the input dimensionality is about 15 times larger than MNIST digit.", "As another example, predicting binary attributes (i.e., 'young', 'bald', 'receding hairline') from human face images (CelebA dataset (Liu et al., 2015) ) ought to be related to the age group classification using human photos taken in the wild (Adience dataset (Eidinger et al., 2014) ).", "The Adience task turns out to be the more difficult one since the wild images are not preprocessed and 7.6 times fewer than CelebA samples.", "Hence, it makes good sense to jointly learn these multi-task representation learning (DMTRL) for CNN setting and our TRMTL (general setting and CNN setting) w.r.t. two tasks.", "The shared portion is depicted in yellow.", "MRN: original weights are totally shared at the lower layers and the relatedness between tasks at the top layers is modelled by tensor normal priors.", "DMTRL (TT or Tucker): all layer-wise weights must be equal-shape so as to be stacked and decomposed into factors.", "For each task, almost all the factors are shard at each layer except the very last 1D vector.", "Such pattern of sharing is identical at all layers.", "TRMTL (General): layer-wise weights are separately encoded into TR-formats for different tasks, and a subset of latent cores are selected to be tied across two tasks.", "The portions of sharing can be different from layer to layer.", "TRMTL (CNN): spatial cores (height and width cores) in the tensorized convolutional kernel are shared, while cores of input/output channels of the kernel are task-specific.", "tasks to extract better feature representations, especially for the hard tasks, which could be achieved through transferring domain-specific knowledge from easy tasks.", "Unfortunately, existing cutting-edge deep MTL models are only suited for the multi-label learning where different tasks share the same training inputs (i.e., X i = X j for i = j, where X i denotes the input for task T i ), and thus cannot be directly applied to above learning scenarios.", "This is due to those models fail to provide knowledge-sharing mechanisms that can cope with the intrinsic discrepancies among network architectures across tasks.", "Such discrepancies either arise from the heterogeneous dimensions of input data or from the heterogeneous designs of layer-wise structures.", "Conventionally, knowledge-sharing mechanisms of deep MTL can be hard or soft parameter sharing (Ruder, 2017) .", "Hard sharing models (Zhang et al., 2014; Yin & Liu, 2017) share all parameters at the lower layers but with no parameters being shared at the upper layers across tasks.", "Soft sharing models (Duong et al., 2015; Yang & Hospedales, 2016; Long & Wang, 2015) , on the other hand, learn one DNN per task with its own set of parameters, and the tasks are implicitly connected through imposing regularization terms on the aligned weights.", "The common issue with above mechanisms is that, for the sharing part, the network architectures of all tasks are strictly required to be identical.", "It turns out that some of the tasks have to compromise on a sub-optimal network architecture, which may lead to the deterioration in the overall performance.", "Ideally, at all potentially shared layers, each task should be capable of encoding both task-specific and task-independent portions of variation.", "To overcome this limitation, we propose a latent-subspace knowledge-sharing mechanism that allows to associate each task with distinct source (domain) of data.", "By utilizing tensor representation, different portions of parameters can be shared via latent cores as common knowledge at distinct layers, so that each task can better convey its private knowledge.", "In this work, we realize our proposed framework via tensor ring (TR) format and refer it as tensor ring multi-task learning (TRMTL), as shown in Figure 1 .", "Our main contributions are twofold: (1) we offer a new distributed knowledge-sharing mechanism that can address the discrepancies of network architectures among tasks.", "Compared to existing deep MTL models that are only for multi-label learning, the joint learning of tasks from multi-datasets (multi-domains) with heterogeneous architectures becomes feasible.", "(2) we provide a TR-based implementation of the proposed framework, which further enhances the performance of deep MTL models in terms of both compactness and expressive power.", "Our general TRMTL framework relies on the manual selection of shared cores, i.e., one need to specify the number of shared cores C at each layer if we choose to share the cores in a left-to-right order across tasks.", "Although we can employ some efficient heuristics, the search space of this hyperparameter may grow rapidly as number of the layers increase.", "Besides the greedy search, a more sophisticated and possible option is to automatically select sharable core pairs that have the highest similarity.", "We may consider two cores as a candidate pair if the same perturbation of the two cores induces similar changes in the errors of respective tasks.", "In this way, one can adaptively select most similar cores from tasks according to a certain threshold, leaving the rest as private cores.", "We should also point out that tensorization operation plays a key role in our proposed sharing mechanism.", "Due to the tensorization, the cores can be shared in a much finer granularity via our TRMTL framework.", "Furthermore, tensorizing weight matrix into high-order weight tensor yields more compact tensor network format (with much lower overall ranks), and thus a higher compression ratio for parameters.", "In contrast, DMTRL tends to produce a lot more parameters without tensorization.", "In this work, we have extended the conventional deep MTL to a broader paradigm where multiple tasks may involve more than one source data domain.", "To resolve the issues caused by the discrepancies among different tasks' network structures, we have introduced a novel knowledge sharing framework for deep MTL, by partially sharing latent cores via tensor network format.", "Our method is empirically verified on various learning settings and achieves the state-of-the-art results in helping tasks to improve their overall performance.", "of T tasks to be equal-sized, so that these weights could be stacked up into one weight matrix W ∈ R M ×T .", "The work (Kumar & Daume III, 2012 ) assumes W to be low-rank and factorizes it as W = LS.", "Here, L ∈ R M ×K consists of K task-independent latent basis vectors, whereas each column vector of S ∈ R K×T is task-specific and contains the mixing coefficients of these common latent bases.", "Yang & Hospedales (2017) extended this to its tensorial counterpart deep multi-task representation learning (DMTRL) by making use of tensor factorization.", "Likewise, DMTRL starts by putting the equal-shaped weight matrices", "side by side along the 'task' mode to form a 3rd-order weight tensor W ∈ R M ×N ×T .", "In the case of CNN, this weight tensor corresponds to a 5th-order filter tensor K ∈ R H×W ×U ×V ×T .", "DMTRL then factorizes W (or K), for instance via TT-format, into 3 TT-cores (or 5 TT-cores for K) Yang & Hospedales (2017) .", "Analogously, the first 2 TT-cores (or the first 4 TT-cores) play exactly the same role as L for the common knowledge; the very last TT-core is in fact a matrix (similar to S), with each column representing the task-specific information.", "The fundamental difference between our TRMTL and DMTRL is that ours can tailor heterogeneous network structures to various tasks.", "In contrast, DMTRL is not flexible enough to deal with such variations with tasks.", "Specifically, our TRMTL differs widely with DMTRL and generalizes DMTRL from a variety of aspects.", "In order to reach TRMTL from DMTRL-TT, one needs to take four major types of generalizations (G1-G4), as shown in Figure 6 .", "Firstly (in G1), TRMTL tensorizes the weight into a higher-order weight tensor before factorizing it.", "By doing so, the weight can be embedded into more latent cores than that of just 3 cores (or 5 cores) in DMTRL, which yields a more compact model and makes the sharing at a finer granularity feasible.", "Secondly (in G2), DMTRL stringently requires that the first D-1 cores (D is weight tensor's order) must be all shared at every hidden layer, only the last vector is kept for private knowledge.", "By contrast, TRMTL allows for any sharing pattern at distinct layer.", "Thirdly (in G3), there is no need for layerwise weights to be equal-sized and stacked into one big tensor as in TRMTL, each task may have its individual input domains.", "Finally (in G4), TRMTL further generalizes TT to TR-format.", "For each task in DMTRL, the first core must be a matrix and the last core must be a vector (with both border rank and outer mode size being 1).", "Notice that our TRMTL also conceptually subsumes DMTRLTucker in terms of the first three aspects of generalizations (G1-G3).", "It is also worth mentioning that (Wang et al., 2018) only applies TR-format for weight compression in a single deep net, whereas ours incorporates a more general tensor network framework into the deep MTL context.", "The authors of (Long et al., 2017 ) lately proposed multilinear relationship network (MRN) which incorporates tensor normal priors over the parameter tensors of the task-specific layers.", "However, like methods (Zhang et al., 2014; Ouyang et al., 2014; Chu et al., 2015) , MRN follows the architecture where all the lower layers are shared, which is also not tailored for the extended MTL paradigm, and may harm the transferability if tasks are not that tightly correlated.", "In addition, the relatedness of tasks is captured by the covariance structures over features, classes and tasks.", "Constantly updating these covariance matrices (via SVD in (Long et al., 2017) ) becomes computationally prohibitive for large scale networks.", "Compared to these non-latent-subspace methods, TRMTL is highly compact and needs much fewer parameters, which is obviously advantageous in tasks with small sample size.", "The detailed specification of network architecture and factorized TRRL representation of the experiments on MNIST dataset are recorded in Table 6 .", "In Table 7 , our TRMTL achieves the best results and is robust to small perturbation of C for pattern selection, since both '410' and '420' patterns obtain similarly good performance." ]

[ 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.1249999925494194, 0.14999999105930328, 0.07407406717538834, 0.3571428656578064, 0.10526315122842789, 0.11764705181121826, 0, 0.10810810327529907, 0.08888888359069824, 0.08510638028383255, 0.08163265138864517, 0, 0.09999999403953552, 0, 0, 0, 0, 0, 0.1666666567325592, 0, 0, 0, 0, 0, 0.11428570747375488, 0, 0, 0.06451612710952759, 0.11538461595773697, 0.0624999962747097, 0, 0.1599999964237213, 0, 0, 0.060606054961681366, 0.060606054961681366, 0, 0.1249999925494194, 0, 0.1764705777168274, 0.1818181723356247, 0.17142856121063232, 0.11764705181121826, 0.08888888359069824, 0, 0.06451612710952759, 0.06451612710952759, 0.0624999962747097, 0.07407406717538834, 0.14814814925193787, 0.11428570747375488, 0.09090908616781235, 0.11428570747375488, 0.20512820780277252, 0.0624999962747097, 0, 0, 0, 0.19354838132858276, 0, 0.0714285671710968, 0.06666666269302368, 0.06896550953388214, 0.09090908616781235, 0, 0, 0.0833333283662796, 0, 0.0833333283662796, 0.045454543083906174, 0.04878048226237297, 0.0952380895614624, 0.04999999701976776, 0, 0.05882352590560913, 0, 0.1818181723356247, 0, 0.04081632196903229, 0, 0.06451612710952759, 0, 0, 0.04999999701976776 ]

[ "a distributed latent-space based knowledge-sharing framework for deep multi-task learning" ]

[ "Board games often rely on visual information such as the location of the game pieces and textual information on cards.", "Due to this reliance on visual feedback, blind players are at a disadvantage because they cannot read the cards or see the location of the game pieces and may be unable to play a game without sighted help.", "We present Game Changer, an augmented workspace that provides both audio descriptions and tactile additions to make the state of the board game accessible to blind and visually impaired players.", "In this paper, we describe the design of Game Changer and present findings from a user study in which 7 blind participants used Game Changer to play against a sighted partner.", "Most players stated the game was more accessible with the additions from Game Changer and felt that Game Changer could be used to augment other games." ]

[ 0, 0, 1, 0, 0 ]

[ 0.1904761791229248, 0.27586206793785095, 0.807692289352417, 0.30188679695129395, 0.4166666567325592 ]

[ "Game Changer is a system that provides both audio descriptions and tactile additions to make the state of the board game accessible to blind and visually impaired players." ]

[ "In many applications labeled data is not readily available, and needs to be collected via pain-staking human supervision.\n", "We propose a rule-exemplar model for collecting human supervision to combine the scalability of rules with the quality of instance labels. ", "The supervision is coupled such that it is both natural for humans and synergistic for learning.", "We propose a training algorithm that jointly denoises rules via latent coverage variables, and trains the model through a soft implication loss over the coverage and label variables. ", "Empirical evaluation on five different tasks shows that (1) our algorithm is more accurate than several existing methods of learning from a mix of clean and noisy supervision, and (2) \nthe coupled rule-exemplar supervision is effective in denoising rules.", "With the ever-increasing reach of machine learning, a common hurdle to new adoptions is the lack of labeled data and the pain-staking process involved in collecting it via human supervision.", "Over the years, several strategies have evolved for reducing the tedium of collecting human supervision.", "On the one hand are methods like active learning and crowd-consensus learning that seek to reduce the cost of supervision in the form of per-instance labels.", "On the other hand is the rich history of rule-based methods (Appelt et al., 1993; Cunningham, 2002) where humans code-up their supervision as labeling rules.", "There is growing interest in learning from such scalable, albiet noisy, supervision (Ratner et al., 2016; Pal & Balasubramanian, 2018; Bach et al., 2019; Sun et al., 2018; Kang et al., 2018 ).", "However, clean task-specific instance labels continue to be critical for reliable results (Goh et al., 2018; Bach et al., 2019) even when fine-tuning models pre-trained on indirect supervision (Sun et al., 2017; Devlin et al., 2018) .", "In this paper we propose a unique blend of cheap coarse-grained supervision in the form of rules and expensive fine-grained supervision in the form of labeled instances.", "Instead of supervising rules and instance labels independently, we propose that each labeling rule be attached with exemplars of where the rule correctly 'fires'.", "Thus, the rule can be treated as a noisy generalization of those exemplars.", "Often rules are coded up only after inspecting data.", "As a human inspects instances, he labels them, and then generalizes them to rules.", "Thus, humans provide paired supervision of rules and exemplars demonstrating correct deployment of that rule.", "We explain further with two illustrative applications.", "Our examples below are from the text domain because rules have been traditionally used in many NLP tasks, but our learning algorithm is agnostic to how rules are expressed.", "Sentiment Classification Consider an instance I highly recommend this modest priced cellular phone that a human inspects for a sentiment labeling task.", "After labeling it as positive, he can easily generalize it to a rule Contains 'highly recommend' → positive label.", "This rule generalizes to several more instances, thereby eliminating the need of per-instance labeling on those.", "However, the label assigned by this rule on unseen instances may not be as reliable as the explicit label on this specific exemplar it generalized.", "For example, it misfires on I would highly recommend this phone if it weren't for their poor service.", "Slot-filling Consider a slot-filling task on restaurant reviews over labels like cuisine, location, and time.", "When an annotator sees an instance like: what chinese restaurants in this city have good reviews?", ", after labeling token chinese as cuisine, he generalizes it to a rule: (. * ese|. * ian|mexican) restaurants → (cuisine) restaurants.", "This rule matches hundreds of instances in the unlabeled set, but could wrongly label a phrase like these restaurants.", "We present in Section 3 other applications where such supervision is natural.", "Our focus in this paper is developing algorithms for training models under such coupled rule-exemplar supervision.", "Our main challenge is that the labels induced by the rules are more noisy than instance-level supervised labels because humans tend to over generalize (Tessler & Goodman, 2019) as we saw in the illustrations above.", "Learning with noisy labels with or without additional clean data has been a problem of long-standing interest in ML (Khetan et al., 2018; Zhang & Sabuncu, 2018; Ren et al., 2018b; Veit et al., 2017; Shen & Sanghavi, 2019) .", "However, we seek to design algorithms that better capture rule-specific noise with the help of exemplars around which we have supervision that the rule fired correctly.", "We associate a latent random variable on whether a rule correctly 'covers' an instance, and jointly learn the distribution among the label and all cover variables.", "This way we simultaneously train the classifier with corrected rule-label examples, and restrict over-generalized rules.", "In summary our contributions in this paper are as follows:", "Our contributions (1) We propose the paradigm of supervision in the form of rules generalizing labeled exemplars that is natural in several applications.", "(2) We design a training method that simultaneously denoises over-generalized rules via latent coverage variables, and trains a classification model with a soft implication loss that we introduce.", "(3) Through experiments on five tasks spanning question classification, spam detection, sequence labeling, and record classification we show that our proposed paradigm of supervision enables an effective synergy between rule-level and instance-level supervision.", "(4) We compare our algorithm to several recent frameworks for learning with noisy supervision and constraints, and show much better results with our method.", "We proposed a new rule-exemplar model for collecting human supervision to combine the scalability of top-level rules with the quality of instance-level labels.", "We show that such supervision is natural since humans typically inspect examples to code rules.", "Furthermore, such coupled examples provide supervision on correct firing of rules which help to denoise rules.", "We propose to train the classifier while jointly denoising rules via latent coverage variables imposing a soft-implication constraint on the true label.", "Empirically on five datasets we show that our training algorithm that performs rule-specific denoising is better than generic noise-tolerant learning.", "In future we plan to deploy this framework on other applications where human supervision is a scarce resource.", "We model a joint distribution Q(y, r 1 , . . . , r n |x) to capture the interaction among the label random variable y and coverage random variables r 1 , . . . , r n of any instance x.", "We use r to compactly represent r 1 , . . . , r n .", "Strictly speaking, when a rule R j does not cover x, the r j is not a random variable and its value is pinned to 0 but we use this fixed-tuple notation for clarity.", "The random variables r j and y impose a constraint on the joint distribution Q: for a x ∈ H j when r j = 1, the label y cannot be anything other than j .", "r j = 1 =⇒ y = j ∀x ∈ H j (7) We can convert this into a soft constraint on the marginals of the distribution Q by stating the probability of y = j Q(y, r j = 1|x) should be small.", "The singleton marginals of Q along the y and r j variables are tied to the P θ and P jφ (r j |x) we seek to learn.", "A network with parameters θ models the classifier P θ (y|x), and a separate network with φ variables (shared across all rules) learns the P jφ (r j |x) distribution.", "The marginals of joint Q should match these trained marginals and we use a KL term for that:", "We call the combined KL term succinctly as KL(Q, P θ ) + KL(Q, P φ ).", "Further the P θ and P jφ distributions should maximize the log-likelihood on their respective labeled data as provided in Equation 1 and Equation 2 respectively.", "Putting all the above objectives together with hyper-parameters α > 0, λ > 0 we get our final objective as:", "We show in Section A.1 that this gives rise to the solution for Q in terms of P θ , P jφ and alternately for P θ , P jφ in terms of Q as follows.", "where δ(y = j ∧ r j = 1) is an indicator function that is 1 when the constraint inside holds, else it is 0.", "Computing marginals of the above using straight-forward message passing techniques we get:", "(13) Thereafter, we solve for θ and φ in terms of a given Q as", "Here, γ = 1 α .", "This gives rise to an alternating optimization algorithm as in the posterior regularization framework of Ganchev et al. (2010) .", "We initialize θ and φ randomly.", "Then in a loop, we perform the following two steps alternatively much like the EM algorithm (Dempster et al., 1977) ." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.1764705777168274, 0.34285715222358704, 0.13793103396892548, 0.29999998211860657, 0.19607843458652496, 0.1904761791229248, 0.06896550953388214, 0.21621620655059814, 0.09999999403953552, 0.0476190447807312, 0.1304347813129425, 0.2222222238779068, 0.1621621549129486, 0.0714285671710968, 0.0833333283662796, 0.3448275923728943, 0.20689654350280762, 0, 0.0952380895614624, 0.0555555522441864, 0.12121211737394333, 0.06451612710952759, 0, 0, 0.19999998807907104, 0, 0.11428570747375488, 0.05882352590560913, 0.07407406717538834, 0.12903225421905518, 0.12765957415103912, 0.0833333283662796, 0.10526315122842789, 0.21052631735801697, 0.13333332538604736, 0, 0.11428570747375488, 0.25, 0.08695651590824127, 0.1666666567325592, 0.3333333432674408, 0.19999998807907104, 0.2666666507720947, 0.2222222238779068, 0, 0.1818181723356247, 0.1395348757505417, 0.07999999821186066, 0.13333332538604736, 0.09302324801683426, 0.04255318641662598, 0.15789473056793213, 0.09999999403953552, 0.1249999925494194, 0, 0.05405404791235924, 0, 0.09999999403953552, 0, 0, 0.13333332538604736, 0, 0.05882352590560913, 0.0952380895614624, 0.05714285373687744 ]

[ "Coupled rule-exemplar supervision and a implication loss helps to jointly learn to denoise rules and imply labels." ]

[ "We consider a problem of learning the reward and policy from expert examples under unknown dynamics.", "Our proposed method builds on the framework of generative adversarial networks and introduces the empowerment-regularized maximum-entropy inverse reinforcement learning to learn near-optimal rewards and policies.", "Empowerment-based regularization prevents the policy from overfitting to expert demonstrations, which advantageously leads to more generalized behaviors that result in learning near-optimal rewards.", "Our method simultaneously learns empowerment through variational information maximization along with the reward and policy under the adversarial learning formulation.", "We evaluate our approach on various high-dimensional complex control tasks.", "We also test our learned rewards in challenging transfer learning problems where training and testing environments are made to be different from each other in terms of dynamics or structure.", "The results show that our proposed method not only learns near-optimal rewards and policies that are matching expert behavior but also performs significantly better than state-of-the-art inverse reinforcement learning algorithms.", "Reinforcement learning (RL) has emerged as a promising tool for solving complex decision-making and control tasks from predefined high-level reward functions BID23 .", "However, defining an optimizable reward function that inculcates the desired behavior can be challenging for many robotic applications, which include learning social-interaction skills BID17 , dexterous manipulation BID5 , and autonomous driving BID10 .Inverse", "reinforcement learning (IRL) BID14 addresses the problem of learning reward functions from expert demonstrations, and it is often considered as a branch of imitation learning BID2 ). The prior", "work in IRL includes maximum-margin BID0 BID18 and maximum-entropy BID24 formulations. Currently", ", maximum entropy (MaxEnt) IRL is a widely used approach towards IRL, and has been extended to use non-linear function approximators such as neural networks in scenarios with unknown dynamics by leveraging sampling-based techniques BID3 BID5 BID9 . However,", "designing the IRL algorithm is usually complicated as it requires, to some extent, hand engineering such as deciding domain-specific regularizers BID5 .Rather than", "learning reward functions and solving the IRL problem, imitation learning (IL) learns a policy directly from expert demonstrations. Prior work", "addressed the IL problem through behavior cloning (BC), which learns a policy from expert trajectories using supervised learning BID15 . Although BC", "methods are simple solutions to IL, these methods require a large amount of data because of compounding errors induced by covariate shift BID19 . To overcome", "BC limitations, a generative adversarial imitation learning (GAIL) algorithm BID8 was proposed. GAIL uses the", "formulation of Generative Adversarial Networks (GANs) BID7 , i.e., a generator-discriminator framework, where a generator is trained to generate expert-like trajectories while a discriminator is trained to distinguish between generated and expert trajectories. Although GAIL", "is highly effective and efficient framework, it does not recover transferable/portable reward functions along with the policies, thus narrowing its use cases to similar problem instances in similar environments. Reward function", "learning is ultimately preferable, if possible, over direct imitation learning as rewards are portable functions that represent the most basic and complete representation of agent intention, and can be re-optimized in new environments and new agents.Reward learning is challenging as there can be many optimal policies explaining a set of demonstrations and many reward functions inducing an optimal policy BID14 BID24 . Recently, an adversarial", "inverse reinforcement learning (AIRL) framework BID6 , an extension of GAIL, was proposed that offers a solution to the former issue by exploiting the maximum entropy IRL method BID24 whereas the latter issue is addressed through learning disentangled reward functions by modeling the reward as a function of state only instead of both state and action. However, AIRL fails to recover", "the ground truth reward when the ground truth reward is a function of both state and action. For example, the reward function", "in any locomotion or ambulation tasks contains a penalty term that discourages actions with large magnitudes. This need for action regularization", "is well known in optimal control literature and limits the use cases of a state-only reward function in most practical real-life applications. A more generalizable and useful approach", "would be to formulate reward as a function of both states and actions, which induces action-driven reward shaping that has been shown to play a vital role in quickly recovering the optimal policies BID13 .In this paper, we propose the empowerment-regularized", "adversarial inverse reinforcement learning (EAIRL) algorithm 1 . Empowerment BID20 ) is a mutual information-based theoretic", "measure, like state-or action-value functions, that assigns a value to a given state to quantify the extent to which an agent can influence its environment. Our method uses variational information maximization BID12", "to learn empowerment in parallel to learning the reward and policy from expert data. Empowerment acts as a regularizer to policy updates to prevent", "overfitting the expert demonstrations, which in practice leads to learning robust rewards. Our experimentation shows that the proposed method recovers not", "only near-optimal policies but also recovers robust, transferable, disentangled, state-action based reward functions that are near-optimal. The results on reward learning also show that EAIRL outperforms", "several state-of-the-art IRL methods by recovering reward functions that leads to optimal, expert-matching behaviors. On policy learning, results demonstrate that policies learned through", "EAIRL perform comparably to GAIL and AIRL with non-disentangled (state-action) reward function but significantly outperform policies learned through AIRL with disentangled reward (state-only) and GAN interpretation of Guided Cost Learning (GAN-GCL) BID4 .", "This section highlights the importance of empowerment-regularized MaxEnt-IRL and modeling rewards as a function of both state and action rather than restricting to state-only formulation on learning rewards and policies from expert demonstrations.In the scalable MaxEnt-IRL framework BID4 BID6 , the normalization term is approximated by importance sampling where the importance-sampler/policy is trained to minimize the KL-divergence from the distribution over expert trajectories.", "However, merely minimizing the divergence between expert demonstrations and policy-generated samples leads to localized policy behavior which hinders learning generalized reward functions.", "In our proposed work, we regularize the policy update with empowerment i.e., we update our policy to reduce the divergence from expert data distribution as well as to maximize the empowerment (Eqn.12).", "The proposed regularization prevents premature convergence to local behavior which leads to robust state-action based rewards learning.", "Furthermore, empowerment quantifies the extent to which an agent can control/influence its environment in the given state.", "Thus the agent takes an action a on observing a state s such that it has maximum control/influence over the environment upon ending up in the future state s .Our", "experimentation also shows the importance of modeling discriminator/reward functions as a function of both state and action in reward and policy learning under GANs framework. The", "re-ward learning results show that state-only rewards (AIRL(s)) does not recover the action dependent terms of the ground-truth reward function that penalizes high torques. Therefore", ", the agent shows aggressive behavior and sometimes flips over after few steps (see the accompanying video), which is also the reason that crippled-ant trained with AIRL's disentangled reward function reaches only the half-way to expert scores as shown in TAB0 . Therefore", ", the reward formulation as a function of both states and actions is crucial to learning action-dependent terms required in most real-world applications, including any autonomous driving, robot locomotion or manipulation task where large torque magnitudes are discouraged or are dangerous. The policy", "learning results further validate the importance of the state-action reward formulation. TAB2 shows", "that methods with state-action reward/discriminator formulation can successfully recover expert-like policies. Hence, our", "empirical results show that it is crucial to model reward/discriminator as a function of state-action as otherwise, adversarial imitation learning fails to learn ground-truth rewards and expert-like policies from expert data.", "We present an approach to adversarial reward and policy learning from expert demonstrations by regularizing the maximum-entropy inverse reinforcement learning through empowerment.", "Our method learns the empowerment through variational information maximization in parallel to learning the reward and policy.", "We show that our policy is trained to imitate the expert behavior as well to maximize the empowerment of the agent over the environment.", "The proposed regularization prevents premature convergence to local behavior and leads to a generalized policy that in turn guides the reward-learning process to recover near-optimal reward.", "We show that our method successfully learns near-optimal rewards, policies, and performs significantly better than state-of-the-art IRL methods in both imitation learning and challenging transfer learning problems.", "The learned rewards are shown to be transferable to environments that are dynamically or structurally different from training environments.In our future work, we plan to extend our method to learn rewards and policies from diverse human/expert demonstrations as the proposed method assumes that a single expert generates the training data.", "Another exciting direction would be to build an algorithm that learns from sub-optimal demonstrations that contains both optimal and non-optimal behaviors." ]

[ 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.29411762952804565, 0.7317073345184326, 0.3499999940395355, 0.2702702581882477, 0, 0.21276594698429108, 0.3829787075519562, 0.14999999105930328, 0.11764705181121826, 0.27272728085517883, 0.13333332538604736, 0.0714285671710968, 0.09999999403953552, 0.3243243098258972, 0.20512819290161133, 0.04878048226237297, 0.12121211737394333, 0.1249999925494194, 0.1249999925494194, 0.1764705777168274, 0.20895521342754364, 0.12121211737394333, 0, 0.09090908616781235, 0.1818181723356247, 0.1818181723356247, 0.1702127605676651, 0.3684210479259491, 0.3684210479259491, 0.14999999105930328, 0.09999999403953552, 0.1304347813129425, 0.2985074520111084, 0.29999998211860657, 0.1818181723356247, 0.1764705777168274, 0.11764705181121826, 0.09302324801683426, 0.1428571343421936, 0.1463414579629898, 0.14035087823867798, 0.1355932205915451, 0.13333332538604736, 0.06451612710952759, 0.3404255211353302, 0.5128204822540283, 0.3529411852359772, 0.15789473056793213, 0.1904761791229248, 0.1860465109348297, 0.35087719559669495, 0.21052631735801697 ]

[ "Our method introduces the empowerment-regularized maximum-entropy inverse reinforcement learning to learn near-optimal rewards and policies from expert demonstrations." ]

[ "Recurrent neural networks (RNNs) can learn continuous vector representations of symbolic structures such as sequences and sentences; these representations often exhibit linear regularities (analogies).", " Such regularities motivate our hypothesis that RNNs that show such regularities implicitly compile symbolic structures into tensor product representations (TPRs; Smolensky, 1990), which additively combine tensor products of vectors representing roles (e.g., sequence positions) and vectors representing fillers (e.g., particular words)", ". To test this hypothesis, we introduce Tensor Product Decomposition Networks (TPDNs), which use TPRs to approximate existing vector representations", ". We demonstrate using synthetic data that TPDNs can successfully approximate linear and tree-based RNN autoencoder representations, suggesting that these representations exhibit interpretable compositional structure; we explore the settings that lead RNNs to induce such structure-sensitive representations", ". By contrast, further TPDN experiments show that the representations of four models trained to encode naturally-occurring sentences can be largely approximated with a bag of words, with only marginal improvements from more sophisticated structures", ". We conclude that TPDNs provide a powerful method for interpreting vector representations, and that standard RNNs can induce compositional sequence representations that are remarkably well approximated byTPRs; at the same time, existing training tasks for sentence representation learning may not be sufficient for inducing robust structural representations", "Compositional symbolic representations are widely held to be necessary for intelligence BID8 Fodor & Pylyshyn, 1988) , particularly in the domain of language BID7 .", "However, neural networks have shown great success in natural language processing despite using continuous vector representations rather than explicit symbolic structures.", "How can these continuous representations yield such success in a domain traditionally believed to require symbol manipulation?One", "possible answer is that neural network representations implicitly encode compositional structure. This", "hypothesis is supported by the spatial relationships between such vector representations, which have been argued to display geometric regularities that parallel plausible symbolic structures of the elements being represented (Mikolov et al. 2013 ; see Figure 1 ).Analogical", "relationships such as those in Figure 1 are special cases of linearity properties shared by several methods developed in the 1990s for designing compositional vector embeddings of symbolic structures. The most general", "of these is tensor product representations (TPRs; BID22 . Symbolic structures", "are first decomposed into filler-role bindings; for example, to represent the sequence [5, 2, 4] , the filler 5 may be bound to the role of first element, the filler 2 may be bound to the role of second element, and so on. Each filler f i and", "-crucially -each role r i has a vector embedding; these two vectors are combined using their tensor product f i ⊗ r i , and these tensor products are summed to produce the representation of the sequence: f i ⊗ r i . This linear combination", "can predict the linear relations between sequence representations illustrated in Figure 1 . (a) (b) (c) Figure 1 :", "Plots", "of", "the first two principal components of (a) word embeddings BID14", ", (b) digit-sequence embeddings", "learned by an autoencoder (Section 2), and (c) sentences (InferSent: Conneau", "et al. 2017) . All demonstrate systematicity in", "the learned vector spaces.In this article, we test the hypothesis that vector representations of sequences can be approximated as a sum of filler-role bindings, as in TPRs. We introduce the Tensor Product", "Decomposition Network (TPDN) which takes a set of continuous vector representations to be analyzed and learns filler and role embeddings that best predict those vectors, given a particular hypothesis for the relevant set of roles (e.g., sequence indexes or structural positions in a parse tree).To derive structure-sensitive representations", ", in Section 2 we look at a task driven by structure, not content: autoencoding of sequences of meaningless symbols, denoted by digits. The focus here is on sequential structure, although", "we also devise a version of the task that uses tree structure. For the representations learned by these autoencoders", ", TPDNs find excellent approximations that are TPRs.In Section 3, we turn to sentence-embedding models from the contemporary literature.It is an open question how structure-sensitive these representations are; to the degree that they are structuresensitive, our hypothesis is that they can be approximated by TPRs. Here, TPDNs find less accurate approximations, but they", "also show that a TPR equivalent to a bag-of-words already provides a reasonable approximation; these results suggest that these sentence representations are not robustly structure-sensitive. We therefore return to synthetic data in Section 4, exploring", "which architectures and training tasks are likely to lead RNNs to induce structure-sensitive representations.To summarize the contributions of this work, TPDNs provide a powerful method for interpreting vector representations, shedding light on hard-to-understand neural architectures. We show that standard RNNs can induce compositional representations", "that are remarkably well approximated by TPRs and that the nature of these representations depends, in intrepretable ways, on the architecture and training task. Combined with our finding that standard sentence encoders do not seem", "to learn robust representations of structure, these findings suggest that more structured architectures or more structure-dependent training tasks could improve the compositional capabilities of existing models.", "What kind of internal representations could allow simple sequence-to-sequence models to perform the remarkable feats they do, including tasks previously thought to require compositional, symbolic representations (e.g., translation)?", "Our experiments show that, in heavily structure-sensitive tasks, sequence-to-sequence models learn representations that are extremely well approximated by tensorproduct representations (TPRs), distributed embeddings of symbol structures that enable powerful symbolic computation to be performed with neural operations BID23 .", "We demonstrated this by approximating learned representations via TPRs using the proposed tensor-product decomposition network (TPDN).", "Variations in architecture and task were shown to induce different types and degrees of structure-sensitivity in representations, with the decoder playing a greater role than the encoder in determining the structure of the learned representation.", "TPDNs applied to mainstream sentence-embedding models reveal that unstructured bag-of-words models provide a respectable approximation; nonetheless, this experiment also provides evidence for a moderate degree of structuresensitivity.", "The presence of structure-sensitivity is corroborated by targeted analogy tests motivated by the linearity of TPRs.", "A limitation of the current TPDN architecture is that it requires a hypothesis about the representations to be selected in advance.", "A fruitful future research direction would be to automatically explore hypotheses about the nature of the TPR encoded by a network." ]

[ 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.19999998807907104, 0.2222222238779068, 0, 0.1599999964237213, 0.08163265138864517, 0.20338982343673706, 0.1463414579629898, 0.21052631735801697, 0.1764705777168274, 0.06896550953388214, 0.1090909019112587, 0.1702127605676651, 0.07407406717538834, 0.08163265138864517, 0.07999999821186066, 0.060606054961681366, 0, 0, 0.0714285671710968, 0.0833333283662796, 0.08888888359069824, 0.16129031777381897, 0.08888888359069824, 0.05714285373687744, 0, 0.08510638028383255, 0.21052631735801697, 0.0833333283662796, 0, 0.04444443807005882, 0.11320754140615463, 0.060606054961681366, 0.17777776718139648, 0.0952380895614624, 0, 0.10810810327529907, 0.05405404791235924 ]

[ "RNNs implicitly implement tensor-product representations, a principled and interpretable method for representing symbolic structures in continuous space." ]

[ "We address the problem of teaching an RNN to approximate list-processing algorithms given a small number of input-output training examples.", "Our approach is to generalize the idea of parametricity from programming language theory to formulate a semantic property that distinguishes common algorithms from arbitrary non-algorithmic functions.", "This characterization leads naturally to a learned data augmentation scheme that encourages RNNs to learn algorithmic behavior and enables small-sample learning in a variety of list-processing tasks.", "Since the earliest days of neural network research, some of the most important questions about neural models have focused on their ability to capture the crispness, systematicity and compositionality that characterize symbolic computation and human cognition BID2 BID11 , and to do so with a human-like number of examples BID10 .", "While recent studies have demonstrated promising results in training recurrent neural networks (RNNs) to approximate symbolic algorithms in domains like list manipulation BID4 BID7 , binary arithmetic BID8 , graph traversal BID3 , and planar geometry BID12 , the question of sample efficiency remains very much open.", "Difficult algorithmic problems may require tens or hundreds of thousands of labelled training examples, and even simple tasks on small inputs seem to require more data than should be necessary BID9 .Our", "goal in this paper is to teach RNNs to approximate list-processing algorithms f :: DISPLAYFORM0 . Inspired", "by the idea of parametricity BID13 ) from type theory and functional programming, we hypothesize that a feature that distinguishes many algorithms from arbitrary functions is that they commute with some family of element-wise changes to their inputs. We describe", "a method for learning this family from the training set D, and show how this learned information can be used to create an augmented training set for an RNN. Our experiments", "show that this augmentation scheme makes it possible to approximate algorithms from small training sets, in some cases requiring only a single example per input list length." ]

[ 0, 0, 1, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.17142856121063232, 0.14999999105930328, 0.2926829159259796, 0.06896550953388214, 0.03389830142259598, 0.043478257954120636, 0.19354838132858276, 0.11538460850715637, 0.0476190410554409, 0.1860465109348297 ]

[ "Learned data augmentation instills algorithm-favoring inductive biases that let RNNs learn list-processing algorithms from fewer examples." ]

[ "The goal of multi-label learning (MLL) is to associate a given instance with its relevant labels from a set of concepts.", "Previous works of MLL mainly focused on the setting where the concept set is assumed to be fixed, while many real-world applications require introducing new concepts into the set to meet new demands.", "One common need is to refine the original coarse concepts and split them into finer-grained ones, where the refinement process typically begins with limited labeled data for the finer-grained concepts.", "To address the need, we propose a special weakly supervised MLL problem that not only focuses on the situation of limited fine-grained supervision but also leverages the hierarchical relationship between the coarse concepts and the fine-grained ones.", "The problem can be reduced to a multi-label version of negative-unlabeled learning problem using the hierarchical relationship.", "We tackle the reduced problem with a meta-learning approach that learns to assign pseudo-labels to the unlabeled entries.", "Experimental results demonstrate that our proposed method is able to assign accurate pseudo-labels, and in turn achieves superior classification performance when compared with other existing methods.", "Multi-label learning (MLL) is an important learning problem with a wide range of applications BID2 BID0 BID11 .", "While traditional setting focuses on the scenario where the label classes are fixed before learning, many real-world applications face different situations.", "One scenario that is common in many applications is the growing number of classes BID13 , where the growth splits high-level concepts to finer-grained ones BID1 .", "For example, the set of classes might start from high-level concepts such as {Animal, .", ". ., Food }, and then grow to include finer-grained concepts like {Cat, . . ., Dog, . . ., Apple, . . ., Banana}. Typical applications may have collected sufficient number of labeled data for learning the high-level concepts in a fully supervised manner, but it can be challenging for the applications to efficiently adapt the classifier from the high-level (coarse-grained) concepts to the finer-grained ones. Conquering the challenge calls for two components: one is a strategic algorithm to actively collect a few fine-grained and informative labels, and the other is an effective learning model to exploit the fine-grained labels that have been partially collected.This work focuses on the design of the second component-learning an accurate fine-grained classifier with only limited supervision. In particular, we assume that the model receives a data set that contains all the coarse-grained labels and a few fine-grained ones, as shown in FIG0 . Then, the problem of constructing a predictive fine-grained model with the presented data set falls under the big umbrella of weakly supervised learning. Specifically, when we focus on leveraging the coarse-grained labels to build a fine-grained classifier, the problem resembles learning with inexact supervision considered by BID12 , where the coarse-grained labels are not in the exact form for the desired output and could only provide weak information about the target fine-grained labels. On the other hand, if we focus on using the fine-grained part of the labels to train the classifier, the problem can be viewed as a multi-label variant of learning with incomplete supervision as some instances receive their exact fine-grained ground-truth labels whereas some do not have labels at all BID12 . While both the aforementioned problems have attracted much research attention, the combination of them (inexact and incomplete supervision) which our problem of interest can be cast as, has not yet been carefully investigated to the best of our knowledge.Organization In this work, we start from a formal definition of our problem of interest. We then demonstrate a simple way to reduce the original problem into a special form of negative-unlabeled learning problem BID7 leveraging the label hierarchy. To tackle the reduced problem, we begin with a discussion on the caveats carried by some possible existing approaches, and propose a new model that undertakes the challenges posed by inexact and incomplete supervision through a novel learning to learn method which jointly exploits the hierarchical relationship between the coarse-and fine-grained labels, as well as the benefits of all available data in hand. The key idea within our model is to take into account all available information to learn the labeling assignments for the unlabeled entries, called pseudo-labels, and use them to guide the decent direction of the parameter updates on the underlying classifier. Finally, we experimentally demonstrate that the proposed method not only assigns accurate pseudo-labels to the unknown entries but also enjoys significantly better performance than other methods for learning fine-grained classifiers under the limited supervision setting.", "We design a tailored method through a meta-learning strategy, which learns to accurately assign pseudo-labels to the unknown entries of a special weakly supervised MLL problem.", "Experimental results show that our proposed method not only assigns accurate pseudo-labels, but also enable the underlying classifier learned to perform better than other possible existing solutions." ]

[ 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.24390242993831635, 0.07999999821186066, 0.1249999925494194, 0.2222222238779068, 0.31578946113586426, 0.5263158082962036, 0.1666666567325592, 0.21052631735801697, 0.0476190410554409, 0.1304347813129425, 0.0555555522441864, 0.11636363714933395, 0.4444444477558136, 0.20408162474632263 ]

[ "We propose a special weakly-supervised multi-label learning problem along with a newly tailored algorithm that learns the underlying classifier by learning to assign pseudo-labels." ]

[ "We develop a reinforcement learning based search assistant which can assist users through a set of actions and sequence of interactions to enable them realize their intent.", "Our approach caters to subjective search where the user is seeking digital assets such as images which is fundamentally different from the tasks which have objective and limited search modalities.", "Labeled conversational data is generally not available in such search tasks and training the agent through human interactions can be time consuming.", "We propose a stochastic virtual user which impersonates a real user and can be used to sample user behavior efficiently to train the agent which accelerates the bootstrapping of the agent.", "We develop A3C algorithm based context preserving architecture which enables the agent to provide contextual assistance to the user.", "We compare the A3C agent with Q-learning and evaluate its performance on average rewards and state values it obtains with the virtual user in validation episodes.", "Our experiments show that the agent learns to achieve higher rewards and better states.", "Within the domain of search, the recent advances have focused on personalizing the search results through recommendations BID27 BID19 .", "While the quality of recommendations have improved, the conventional search interface has not innovated much to incorporate useful contextual cues which are often missed.", "Conventional search interface enables the end user to perform a keyword based faceted search where the typical work flow goes as follows: the end user types in her search query, applies some filters and then modifies the query based on the results.", "This iterative interaction naturally paves way for incorporating conversations in the process.", "Instead of the search engine just retrieving the best result set, it can interact with the user to collect more contextual cues.", "For example, if a user searches for birthday gift, the search engine could follow-up by asking who are you buying the gift for.", "Such information and interaction can provide more humanlike and engaging search experience along with assisting user in discovering their search intent.", "In this work we address this problem by developing a Reinforcement Learning (RL) BID21 based conversational search agent which interacts with the users to help them in narrowing down to relevant search results by providing them contextual assistance.RL based dialogue agents have been designed for tasks like restaurant, bus and hotel reservation BID18 which have limited and well-defined objective search modalities without much scope for subjective discussion.", "For instance, when searching for a restaurant, the user can specify her preferences (budget, distance, cuisines etc) due to which the problem can be modeled as a slot filling exercise.", "In contrast, suppose a designer is searching for digital assets (over a repository of images, videos etc) to be used in a movie poster.", "She would start with a broad idea and her idea would get refined as the search progresses.", "The modified search intent involves an implicit cognitive feedback which can be used to improve the search results.", "We model our agent for this type of search task.", "Since the user preferences can not be modeled using a fixed set of facets, we end up with a very large search space which is not the case with most other goal oriented RL agents.We model the search process as a sequence of alternate interactions between the user and the RL agent.", "The extent to which the RL agent could help the user depends on the sequence and the type of actions it takes according to user behavior.", "Under the RL framework, intermediate rewards is given to the agent at each step based on its actions and state of conversational search.", "It learns Since true conversational data is not easily available in search domain, we propose to use query and session log data to develop a stochastic virtual user environment to simulate training episodes and bootstrap the learning of the agent.", "Our agent interacts with the user to gauge user intent and treats the search engine as a black box service which makes it easily deployable over any search engine.", "We perform qualitative experiments by simulating validation episodes with different reinforcement learning algorithms under various formulations of the state space to evaluate the performance of the trained agent.Our contributions are three-fold:", "1) formulating conversational interactive search as a reinforcement learning problem and proposing a generic and easily extendable set of states, actions and rewards;", "2) developing a stochastic user model which can be used to efficiently sample user actions while simulating an episode;", "3) we develop A3C (Asynchronous Advantage Actor-Critic) BID15 algorithm based architecture to predict the policy and state value functions of RL agent and compare it with other RL algorithms over performance on validation episodes.", "In this paper, we develop a Reinforcement Learning based search assistant to interact with customers to help them search digital assets suited to their use-case.", "We model the rewards, state space, action space and develop an A3C based architecture which leverages the context of search to predict the policy.", "The trained agent is able to obtain higher average rewards in the validation episodes with virtual user and observes states with better values indicative of providing better search experience.", "We also propose a virtual stochastic user model to interact and train the RL agent in absence of labeled conversational data which accelerates the process of obtaining a bootstrapped agent.As the next step, we would deploy our system to collect true conversational data which can be used to fine tune the current model as well as to train a new model which can generate the natural language responses in addition to deciding the action.", "In different search domains, designing the state and action space can take significant time which makes every situation an absolutely new task to be solved.", "To approach this issue as a future work, another system can be designed which helps in the automation of state space characterization with the help of system query logs." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0 ]

[ 0.19512194395065308, 0.1904761791229248, 0.15789473056793213, 0.05128204822540283, 0.24242423474788666, 0.05128204822540283, 0, 0.060606054961681366, 0.1538461446762085, 0.12244897335767746, 0.0714285671710968, 0.1111111044883728, 0.05405404791235924, 0.11428570747375488, 0.27397260069847107, 0.04651162400841713, 0.10526315122842789, 0.06451612710952759, 0.12121211737394333, 0.07692307233810425, 0.07017543166875839, 0.05405404791235924, 0.15789473056793213, 0.11764705181121826, 0.09756097197532654, 0, 0.1111111044883728, 0.05882352590560913, 0.0416666604578495, 0.31578946113586426, 0.15789473056793213, 0.09302324801683426, 0.0882352888584137, 0.09756097197532654, 0.0952380895614624 ]

[ "A Reinforcement Learning based conversational search assistant which provides contextual assistance in subjective search (like digital assets)." ]

[ "We present a simple approach based on pixel-wise nearest neighbors to understand and interpret the functioning of state-of-the-art neural networks for pixel-level tasks.", "We aim to understand and uncover the synthesis/prediction mechanisms of state-of-the-art convolutional neural networks.", "To this end, we primarily analyze the synthesis process of generative models and the prediction mechanism of discriminative models.", "The main hypothesis of this work is that convolutional neural networks for pixel-level tasks learn a fast compositional nearest neighbor synthesis/prediction function.", "Our experiments on semantic segmentation and image-to-image translation show qualitative and quantitative evidence supporting this hypothesis.", "Convolutional neural networks (CNNs) have revolutionized computer vision, producing impressive results for discriminative tasks such as image classification and semantic segmentation.", "More recently, they have also produced startlingly impressive results for image generation through generative models.", "However, in both cases, such feed-forward networks largely operate as \"black boxes.\"", "As a community, we are still not able to succinctly state why and how such feed-forward functions generate a particular output from a given input.", "If a network fails on a particular input, why?", "How will a network behave on never-before-seen data?", "To answer such questions, there is a renewed interest in so-called explainable AI", "In this paper, we have presented a simple approach based on pixel-wise nearest neighbors to understand and interpret the functioning of convolutional neural networks for spatial prediction tasks.", "Our analysis suggests that CNNs behave as compositional nearest neighbor operators over a training set of patch-label pairs that act as an associative memory.", "But beyond simply memorizing, CNNs can generalize to novel data by composing together local patches from different training instances.", "Also, we argued that networks for pixel-level tasks learn sufficient statistics that enable the gener- Table 1 : We compare compositional nearest neighbors (CompNN) to the baseline CNN and different global nearest neighbor approaches, obtained by matching feature maps from different layers (Global-Bottleneck and Global-Decode2).", "We report mean pixel accuracy and intersection-overunion, where predicted segmentation labels are compared to ground-truth labels.", "We specifically use the embedding learned by BID27 for Facades-to-Labels (Facades) and CityScape, and embedding learned by BID2 for CamVid.", "On average, CompNN performs 5% worse than the baseline CNN, though in some cases (CityScapes) it performs equally.", "However, compositional matching dramatically outperforms global matching, sometimes by a factor of 2X (Facade and CityScape IoU).", "In terms of global matching, the last feature layer (Decode2) strictly outperforms the intermediate Bottleneck layer, but is significantly larger (128 3 versus 512 dimensions).", "Finally, self-supervised labels (SS) overall perform similarly to the original labels (O), but almost consistently help for compositional matching and consistently hurt for global matching.", "We posit that this is due to the fact that self-supervised labels tend to be overly-smoothed, and so act as a form of spatial regularization for compositional matching.", "ation of pixel predictions.", "Our analysis and experiments not only support this argument, but also enables example-based explanations of network behavior and explicit modulation of the implicit biases learned by the network.", "We hope that our framework enables further analysis of convolutional networks from a non-parametric perspective.", "FIG0 : Global NN v.s. Comp NN.", "We show synthesized images using our CompNN methods and four global NN approaches (global nearest neighbor on bottleneck feature embedding and Decode2 feature embedding using self-supervised labels and original labels respectively).", "We can observe that (1) compositional nearest neighbor outperforms other global nearest neighbor approaches, (2) using Decode2 features (the penultimate layer) sometimes can generate more similar structures (See row 1,4).", "FIG0 shows the synthesized images using several global NN approaches and a CompNN approach.", "We can observe that the results of global NN approaches overall resembles global properties of the output of the Convolutional Neural Network (CNN) and of the CompNN approach.", "For instance, in the top two rows, the output of the global NN resembles the color of the facade and structural properties of the buildings.", "Also, in the bottom two rows, we can observe that the global NN overall captures the organization of the scene because many labels in the global NN overlap considerably with the output of the CNN and the ground truth." ]

[ 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0, 0, 0, 0, 0, 0.13793103396892548, 0, 0.0952380895614624, 0, 0, 0.1249999925494194, 0, 0, 0.13333332538604736, 0, 0, 0, 0, 0, 0, 0, 0, 0.05882352590560913, 0, 0, 0, 0, 0, 0, 0, 0.13793103396892548, 0, 0 ]

[ "Convolutional Neural Networks behave as Compositional Nearest Neighbors!" ]

[ "Consider a world in which events occur that involve various entities.", "Learning how to predict future events from patterns of past events becomes more difficult as we consider more types of events.", "Many of the patterns detected in the dataset by an ordinary LSTM will be spurious since the number of potential pairwise correlations, for example, grows quadratically with the number of events.", "We propose a type of factorial LSTM architecture where different blocks of LSTM cells are responsible for capturing different aspects of the world state.", "We use Datalog rules to specify how to derive the LSTM structure from a database of facts about the entities in the world.", "This is analogous to how a probabilistic relational model (Getoor & Taskar, 2007) specifies a recipe for deriving a graphical model structure from a database.", "In both cases, the goal is to obtain useful inductive biases by encoding informed independence assumptions into the model.", "We specifically consider the neural Hawkes process, which uses an LSTM to modulate the rate of instantaneous events in continuous time.", "In both synthetic and real-world domains, we show that we obtain better generalization by using appropriate factorial designs specified by simple Datalog programs.\n", "Temporal sequence data is abundant in applied machine learning.", "A common task is to impute missing events, e.g., to predict the future from the past.", "Often this is done by fitting a generative probability model.", "For evenly spaced sequences, historically popular models have included hidden Markov models and discrete-time linear dynamical systems, with more recent interest in recurrent neural network models such as LSTMs.", "For irregularly spaced sequences, a good starting point is the Hawkes process, a self-exciting temporal point process; many variations and enhancements have been published, including neural variants using LSTMs.", "All of these models can be described schematically by Figure 1a .", "Events e i , e i+1 , . . . are assumed to be conditionally independent of previous events, given the system state s i (which may or may not be fully known given events e 1 , . . . , e i ).", "That is, s i is enough to determine the joint distribution of the i th event and the updated state s i+1 , which is needed to recursively predict all subsequent events.", "Figure 1a and its caption show the three types of influence in the model.", "The update, affect, and depend arrows are characterized by parameters of the model.", "In the case of a recurrent neural network, these are the transition, input, and output matrices.", "Our main idea in this paper is to inject structural zeros into these weight matrices.", "Structural zeros are weights that are fixed at zero regardless of the model parameters.", "In other words, we will remove many connections (synapses) from both the recurrent and non-recurrent portions of the neural network.", "Parameter estimation must use the sparse remaining connections to explain the observed data.", "Specifically, we partition the neural state s i ∈ R d into a number of node blocks.", "Different node blocks are intended to capture different aspects of the world's state at step i.", "By zeroing out rectangular blocks of the weight matrix, we will restrict how these node blocks interact with the events and with one another.", "An example is depicted in Figures 1b (affect, depend) and 1d (update).", "In addition, by reusing nonzero blocks within a weight matrix, we can stipulate (for example) that event e affects node block b in the same way in which event e affects node block b .", "Such parameter tying makes it possible to generalize from frequent events to rare events of the same type.", "Although our present experiments are small, we are motivated by the challenges of scale.", "Real-world domains may have millions of event types, including many rare types.", "To model organizational behavior, we might consider a dataset of meetings and emails in a large organization.", "To model supply chains, we might consider purchases of goods and services around the world.", "In an unrestricted model, anything in the past could potentially influence anything in the future, making estimation extremely difficult.", "Structural zeroes and parameter tying, if chosen carefully, should help us avoid overfitting to coincidental patterns in the data.", "Analogous architectures have been proposed in the world of graphical models and causal models.", "Indeed, to write down such a model is to explicitly allow specific direct interactions and forbid the rest.", "For example, the edges of a Gaussian graphical model explicitly indicate which blocks of the inverse covariance matrix are allowed to be nonzero.", "Some such models reuse blocks (Hojsgaard & Lauritzen, 2008) .", "As another example, a factorial HMM (Ghahramani & Jordan, 1997 )-an HMM whose states are m-tuples-can be regarded as a simple example of our architecture.", "The state s i can be represented using m node blocks, each of which is a 1-hot vector that encodes the value of a different tuple element.", "The key aspect of a factorial HMM is that the stochastic transition matrix (update in Figure 1d ) is fully block-diagonal.", "The affect matrix is 0, since the HMM graphical model does not feed the output back into the next state; the depend matrix is unrestricted.", "But how do we know which interactions to allow and which to forbid?", "This is a domain-specific modeling question.", "In general, we would like to exploit the observation that events are structured objects with participants (which is why the number of possible event types is often large).", "For example, a travel event involves both a person and a place.", "We might assume that the probability that Alice travels to Chicago depends only on Alice's state, the states of Alice's family members, and even the state of affairs in Chicago.", "Given that modeling assumption, parameter estimation cannot try to derive this probability (presumably incorrectly) from the state of the coal market.", "These kinds of systematic dependencies can be elegantly written down using Datalog rules, as we will show.", "Datalog rules can refer to database facts, such as the fact that Alice is a person and that she is related to other people.", "Given these facts, we use Datalog rules to automatically generate the set of possible events and node blocks, and the ways in which they influence one another.", "Datalog makes it easy to give structured names to the events and node blocks.", "The rules can inspect these structures via pattern-matching.", "In short, our contribution is to show how to use a Datalog program to systematically derive a constrained neural architecture from a database.", "Datalog is a blend of logic and databases, both of which have previously been used in various formalisms for deriving a graphical model architecture from a database (Getoor & Taskar, 2007) .", "There has been extensive research about having inductive biases in the architecture design of a machine learning model.", "The epitome of this direction is perhaps the graphical models where edges between variables are usually explicitly allowed or forbidden (Koller & Friedman, 2009 ).", "There has also been work in learning such biases from data.", "For example, Stepleton et al. (2009) proposed to encourage the block-structured states for Hidden Markov Models (HMM) by enforcing a sparsityinducing prior over the non-parametric Bayesian model.", "Duvenaud et al. (2013) and Bratières et al. (2014) attempted to learn structured kernels for Gaussian processes.", "Our work is in the direction of injecting inductive biases into a neural temporal model-a class of models that is useful in various domains such as demand forecasting (Seeger et al., 2016) , personalization and recommendation (Jing & Smola, 2017) , event prediction (Du et al., 2016) and knowledge graph modeling (Trivedi et al., 2017) .", "Incorporating structural knowledge in the architecture design of such a model has drawn increasing attention over the past few years.", "Shelton & Ciardo (2014) introduced a factored state space in continuous-time Markov processes.", "Meek (2014) and Bhattacharjya et al. (2018) proposed to consider direct dependencies among events in graphical event models.", "Wang et al. (2019) developed a hybrid model that decomposes exchangeable sequences into a global part that is associated with common patterns and a local part that reflects individual characteristics.", "However, their approaches are all bounded to the kinds of inductive biases that are easy to specify (e.g. by hand).", "Our work enables people to use a Datalog program to conveniently specify the neural architecture based on a deductive database-a much richer class of knowledge than the previous work could handle.", "Although logic programming languages and databases have both previously been used to derive a graphical model architecture (Getoor & Taskar, 2007) , we are, to the best of our knowledge, the first to develop such a general interface for a neural event model.", "As future work, we hope to develop an extension where events can also trigger assertions and retractions of facts in the Datalog database.", "Thanks to the Datalog rules, the model architecture will dynamically change along with the facts.", "For example, if Yoyodyne Corp. hires Alice, then the Yoyodyne node block begins to influence Alice's actions, and K expands to include a new (previously impossible) event where Yoyodyne fires Alice.", "Moreover, propositions in the database-including those derived via other Datalog rules-can now serve as extra bits of system state that help define the λ k intensity functions in (1).", "Then the system's learned neural state s i is usefully augmented by a large, exact set of boolean propositions-a division of labor between learning and expert knowledge.", "In this section, we elaborate on the details of the transition function Ψ that is introduced in section 2.1; more details about them may be found in Mei & Eisner (2017) .", "where the interval (t i−1 , t i ] has consecutive observations k i−1 @t i−1 and k i @t i as endpoints.", "At t i , the continuous-time LSTM reads k i @t i and updates the current (decayed) hidden cells c(t) to new initial values c i+1 , based on the current (decayed) hidden state h(t i ), as follows:", "At time t i , the updated state vector is", "] is given by (26), which continues to control h(t) except that i has now increased by 1).", "On the interval (t i , t i+1 ], c(t) follows an exponential curve that begins at c i+1 (in the sense that lim t→t + i c(t) = c i+1 ) and decays, as time t increases, toward c i+1 (which it would approach as t → ∞, if extrapolated)." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0, 0.060606054961681366, 0.09756097197532654, 0.0555555522441864, 0.1666666567325592, 0.05405404791235924, 0.1764705777168274, 0.1111111044883728, 0.2631579041481018, 0, 0.0624999962747097, 0.07692307233810425, 0.04651162400841713, 0.04651162400841713, 0.07407406717538834, 0.04444443807005882, 0.0952380895614624, 0.06896550953388214, 0.13793103396892548, 0.12903225421905518, 0.25806450843811035, 0, 0.05714285373687744, 0.0714285671710968, 0, 0.0624999962747097, 0.10810810327529907, 0.0714285671710968, 0.09302324801683426, 0.0624999962747097, 0.06896550953388214, 0, 0.0624999962747097, 0.06451612710952759, 0, 0.11428570747375488, 0.06896550953388214, 0.12121211737394333, 0.05405404791235924, 0, 0.05128204822540283, 0, 0, 0, 0.14814814925193787, 0, 0.0476190410554409, 0.07692307233810425, 0.14999999105930328, 0.0555555522441864, 0.060606054961681366, 0.1621621549129486, 0.1463414579629898, 0.20689654350280762, 0, 0.11428570747375488, 0.09090908616781235, 0.05882352590560913, 0, 0.07407406717538834, 0.1428571343421936, 0.12903225421905518, 0.06666666269302368, 0.05714285373687744, 0, 0.11764705181121826, 0.04878048226237297, 0.17142856121063232, 0.09302324801683426, 0.07547169178724289, 0.1538461446762085, 0.13793103396892548, 0.09090908616781235, 0.04651162400841713, 0.0952380895614624, 0, 0.060606054961681366, 0.1304347813129425, 0, 0.12121211737394333, 0.0363636314868927 ]

[ "Factorize LSTM states and zero-out/tie LSTM weight matrices according to real-world structural biases expressed by Datalog programs." ]

[ "In this work, we aim to solve data-driven optimization problems, where the goal is to find an input that maximizes an unknown score function given access to a dataset of input, score pairs.", "Inputs may lie on extremely thin manifolds in high-dimensional spaces, making the optimization prone to falling-off the manifold.", "Further, evaluating the unknown function may be expensive, so the algorithm should be able to exploit static, offline data.", "We propose model inversion networks (MINs) as an approach to solve such problems.", "Unlike prior work, MINs scale to extremely high-dimensional input spaces and can efficiently leverage offline logged datasets for optimization in both contextual and non-contextual settings.", "We show that MINs can also be extended to the active setting, commonly studied in prior work, via a simple, novel and effective scheme for active data collection.", "Our experiments show that MINs act as powerful optimizers on a range of contextual/non-contextual, static/active problems including optimization over images and protein designs and learning from logged bandit feedback.", "Data-driven optimization problems arise in a range of domains: from protein design (Brookes et al., 2019) to automated aircraft design (Hoburg & Abbeel, 2012) , from the design of robots (Liao et al., 2019) to the design of neural net architectures (Zoph & Le, 2017) and learning from logged feedback, such as optimizing user preferences in recommender systems.", "Such problems require optimizing unknown reward or score functions using previously collected data consisting of pairs of inputs and corresponding score values, without direct access to the score function being optimized.", "This can be especially challenging when valid inputs lie on a low-dimensional manifold in the space of all inputs, e.g., the space of valid aircraft designs or valid images.", "Existing methods to solve such problems often use derivative-free optimization (Snoek et al.) .", "Most of these techniques require active data collection where the unknown function is queried at new inputs.", "However, when function evaluation involves a complex real-world process, such as testing a new aircraft design or evaluating a new protein, such active methods can be very expensive.", "On the other hand, in many cases there is considerable prior data -existing aircraft and protein designs, and advertisements and user click rates, etc.", "-that could be leveraged to solve the optimization problem.", "In this work, our goal is to develop an optimization approach to solve such optimization problems that can (1) readily operate on high-dimensional inputs comprising a narrow, low-dimensional manifold, such as natural images, (2) readily utilize offline static data, and (3) learn with minimal active data collection if needed.", "We can define this problem setting formally as the optimization problem", "where the function f (x) is unknown, and we have access to a dataset D = {(x 1 , y 1 ), . . . , (x N , y N )}, where y i denotes the value f (x i ).", "If no further data collection is possible, we call this the data-driven model-based optimization setting.", "This can also be extended to the contextual setting, where the aim is to optimize the expected score function value across a context distribution.", "That is,", "where π maps contexts c to inputs x, such that the expected score under the context distribution p 0 (c) is optimized.", "As before, f (c, x) is unknown and we have access to a dataset D = {(c i ,", ", where y i is the value of f (c i , x i ).", "Such contextual problems with logged datasets have been studied in the context of contextual bandits Joachims et al., 2018) .", "A simple way to approach these model-based optimization problems is to train a proxy function f θ (x) or f θ (c, x), with parameters θ, to approximate the true score, using the dataset D. However, directly using f θ (x) in place of the true function f (x) in Equation (1) generally works poorly, because the optimizer will quickly find an input x for which f θ (x) outputs an erroneously large value.", "This issue is especially severe when the inputs x lie on a narrow manifold in a high-dimensional space, such as the set of natural images (Zhu et al., 2016) .", "The function f θ (x) is only valid near the training distribution, and can output erroneously large values when queried at points chosen by the optimizer.", "Prior work has sought to addresses this issue by using uncertainty estimation and Bayesian models (Snoek et al., 2015) for f θ (x), as well as active data collection (Snoek et al.) .", "However, explicit uncertainty estimation is difficult when the function f θ (x) is very complex or when x is high-dimensional.", "Instead of learning f θ (x), we propose to learn the inverse function, mapping from values y to corresponding inputs x.", "This inverse mapping is one-to-many, and therefore requires a stochastic mapping, which we can express as f −1 θ (y, z) → x, where z is a random variable.", "We term such models model inversion networks (MINs).", "MINs provide us with a number of desirable properties: they can utilize static datasets, handle high-dimensional input spaces such as images, can handle contextual problems, and can accommodate both static datasets and active data collection.", "We discuss how to design simple active data collection methods for MINs, leverage advances in deep generative modeling (Goodfellow et al.; Brock et al., 2019) , and scale to very high-dimensional input spaces.", "We experimentally demonstrate MINs in a range of settings, showing that they outperform prior methods on high-dimensional input spaces, perform competitively to Bayesian optimization methods on tasks with active data collection and lower-dimensional inputs, and substantially outperform prior methods on contextual optimization from logged data (Swaminathan & Joachims, a) .", "Prior work has usually considered MBO in the active or \"onpolicy\" setting, where the algorithm actively queries data as it learns.", "In this work, we introduced the data-driven MBO problem statement and devised a method to perform optimization in such scenarios.", "This is important in settings where data collection is expensive and where abundant datasets exist, for example, protein design, aircraft design and drug design.", "Further, MINs define a family of algorithms that show promising results on MBO problems on extremely large input spaces.", "While MINs scale to high-dimensional tasks such as model-based optimization over images, and are performant in both contextual and non-contextual settings, we believe there are a number of interesting open questions for future work.", "The interaction between active data collection and reweighting should be investigated in more detail, and poses interesting consequences for MBO, bandits and reinforcement learning.", "Better and more principled inference procedures are also a direction for future work.", "Another avenue is to study various choices of training objectives in MIN optimization.", "In this section, we show that the inference scheme described in Equation 4, Section 3.2 emerges as a deterministic relaxation of the probabilistic inference scheme described below.", "We re-iterate that in Section 3.2, a singleton x * is the output of optimization, however the procedure can be motivated from the perspective of the following probabilistic inference scheme.", "Let p(x|y) denote a stochastic inverse map, and let p f (y|x) be a probabilistic forward map.", "Consider the following optimization problem: arg max", "where p θ (x|y) is the probability distribution induced by the learned inverse map (in our case, this corresponds to the distribution of f −1 θ (y, z) induced due to randomness in z ∼ p 0 (·)), p f (x|y) is the learned forward map, H is Shannon entropy, and D is KL-divergence measure between two distributions.", "In Equation 4, maximization is carried out over the input y to the inverse-map, and the input z which is captured inp in the above optimization problem, i.e. maximization over z in Equation 4 is equivalent to choosingp subject to the choice of singleton/ Dirac-deltap.", "The Lagrangian is given by:", "In order to derive Equation 4, we restrictp to the Dirac-delta distribution generated by querying the learned inverse map f −1 θ at a specific value of z.", "Now note that the first term in the Lagrangian corresponds to maximizing the \"reconstructed\"ŷ similarly to the first term in Equation 4.", "If p f is assumed to be a Gaussian random variable with a fixed variance, then log p f (ŷ|x) = −||ŷ − µ(x)|| Finally, in order to obtain the log p 0 (z) term, note that, D(p(x|y), p θ (x|y)) ≤ D(δ z (·), p 0 (·)) = − log p 0 (z) (by the data processing inequality for KL-divergence).", "Hence, constraining log p 0 (z) instead of the true divergence gives us a lower bound on L. Maximizing this lower bound (which is the same as Equation 4) hence also maximizes the true Lagrangian L." ]

[ 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.26923075318336487, 0.19999998807907104, 0.04999999329447746, 0.3333333432674408, 0.4680851101875305, 0.35999998450279236, 0.19607841968536377, 0.17910447716712952, 0.11764705181121826, 0.12244897335767746, 0.21621620655059814, 0.04999999329447746, 0.12765957415103912, 0.08888888359069824, 0.1875, 0.3235293924808502, 0.1818181723356247, 0.11538460850715637, 0.15789473056793213, 0.13636362552642822, 0.09090908616781235, 0.1428571343421936, 0, 0.0952380895614624, 0.20779220759868622, 0.11764705181121826, 0.0833333283662796, 0.11320754140615463, 0.04999999329447746, 0.09302324801683426, 0.11999999731779099, 0.06451612710952759, 0.30188679695129395, 0.3333333432674408, 0.317460298538208, 0.09302324801683426, 0.2790697515010834, 0.1395348757505417, 0.24390242993831635, 0.3272727131843567, 0.13333332538604736, 0.1111111044883728, 0.1666666567325592, 0.12765957415103912, 0.19999998807907104, 0.10256409645080566, 0.06666666269302368, 0.09090908616781235, 0.1818181723356247, 0, 0.08163265138864517, 0.15789473056793213, 0.0882352888584137, 0.037735845893621445 ]

[ "We propose a novel approach to solve data-driven model-based optimization problems in both passive and active settings that can scale to high-dimensional input spaces." ]

[ "Generating formal-language represented by relational tuples, such as Lisp programs or mathematical expressions, from a natural-language input is an extremely challenging task because it requires to explicitly capture discrete symbolic structural information from the input to generate the output.", "Most state-of-the-art neural sequence models do not explicitly capture such structure information, and thus do not perform well on these tasks.", "In this paper, we propose a new encoder-decoder model based on Tensor Product Representations (TPRs) for Natural- to Formal-language generation, called TP-N2F.", "The encoder of TP-N2F employs TPR 'binding' to encode natural-language symbolic structure in vector space and the decoder uses TPR 'unbinding' to generate a sequence of relational tuples, each consisting of a relation (or operation) and a number of arguments, in symbolic space.", "TP-N2F considerably outperforms LSTM-based Seq2Seq models, creating a new state of the art results on two benchmarks: the MathQA dataset for math problem solving, and the AlgoList dataset for program synthesis.", "Ablation studies show that improvements are mainly attributed to the use of TPRs in both the encoder and decoder to explicitly capture relational structure information for symbolic reasoning.", "When people perform explicit reasoning, they can typically describe the way to the conclusion step by step via relational descriptions.", "There is ample evidence that relational representations are important for human cognition (e.g., (Goldin-Meadow & Gentner, 2003; Forbus et al., 2017; Crouse et al., 2018; Chen & Forbus, 2018; Chen et al., 2019) ).", "Although a rapidly growing number of researchers use deep learning to solve complex symbolic reasoning and language tasks (a recent review is (Gao et al., 2019) ), most existing deep learning models, including sequence models such as LSTMs, do not explicitly capture human-like relational structure information.", "In this paper we propose a novel neural architecture, TP-N2F, to solve natural-to formal-language generation tasks (N2F).", "In the tasks we study, math or programming problems are stated in naturallanguage, and answers are given as programs, sequences of relational representations, to solve the problem.", "TP-N2F encodes the natural-language symbolic structure of the problem in an input vector space, maps this to a vector in an intermediate space, and uses that vector to produce a sequence of output vectors that are decoded as relational structures.", "Both input and output structures are modelled as Tensor Product Representations (TPRs) (Smolensky, 1990) .", "During encoding, NL-input symbolic structures are encoded as vector space embeddings using TPR 'binding' (following Palangi et al. (2018) ); during decoding, symbolic constituents are extracted from structure-embedding output vectors using TPR 'unbinding' (following Huang et al. (2018; ).", "Our contributions in this work are as follows.", "(i) We propose a role-level analysis of N2F tasks.", "(ii) We present a new TP-N2F model which gives a neural-network-level implementation of a model solving the N2F task under the role-level description proposed in", "(i).", "To our knowledge, this is the first model to be proposed which combines both the binding and unbinding operations of TPRs to achieve generation tasks through deep learning.", "(iii) State-of-the-art performance on two recently developed N2F tasks shows that the TP-N2F model has significant structure learning ability on tasks requiring symbolic reasoning through program synthesis.", "In this paper we propose a new scheme for neural-symbolic relational representations and a new architecture, TP-N2F, for formal-language generation from natural-language descriptions.", "To our knowledge, TP-N2F is the first model that combines TPR binding and TPR unbinding in the encoderdecoder fashion.", "TP-N2F achieves the state-of-the-art on two instances of N2F tasks, showing significant structure learning ability.", "The results show that both the TP-N2F encoder and the TP-N2F decoder are important for improving natural-to formal-language generation.", "We believe that the interpretation and symbolic structure encoding of TPRs are a promising direction for future work.", "We also plan to combine large-scale deep learning models such as BERT with TP-N2F to take advantage of structure learning for other generation tasks.", "In this section, we present details of the experiments of TP-N2F on the two datasets.", "We present the implementation of TP-N2F on each dataset.", "The MathQA dataset consists of about 37k math word problems ((80/12/8)% training/dev/testing problems), each with a corresponding list of multi-choice options and an straight-line operation sequence program to solve the problem.", "An example from the dataset is presented in the Appendix A.4.", "In this task, TP-N2F is deployed to generate the operation sequence given the question.", "The generated operations are executed to generate the solution for the given math problem.", "We use the execution script from Amini et al. (2019) to execute the generated operation sequence and compute the multi-choice accuracy for each problem.", "During our experiments we observed that there are about 30% noisy examples (on which the execution script fails to get the correct answer on the ground truth program).", "Therefore, we report both execution accuracy (the final multi-choice answer after running the execution engine) and operation sequence accuracy (where the generated operation sequence must match the ground truth sequence exactly).", "The AlgoLisp dataset (Polosukhin & Skidanov, 2018 ) is a program synthesis dataset, which has 79k/9k/10k training/dev/testing samples.", "Each sample contains a problem description, a corresponding Lisp program tree, and 10 input-output testing pairs.", "We parse the program tree into a straight-line sequence of commands from leaves to root and (as in MathQA) use the symbol # i to indicate the result of the i th command (generated previously by the model).", "A dataset sample with our parsed command sequence is presented in the Appendix A.4.", "AlgoLisp provides an execution script to run the generated program and has three evaluation metrics: accuracy of passing all test cases (Acc), accuracy of passing 50% of test cases (50p-Acc), and accuracy of generating an exactly matched program (M-Acc).", "AlgoLisp has about 10% noise data (where the execution script fails to pass all test cases on the ground truth program), so we report results both on the full test set and the cleaned test set (in which all noisy testing samples are removed).", "We use d R , n R , d F , n F to indicate the TP-N2F encoder hyperparameters, the dimension of role vectors, the number of roles, the dimension of filler vectors and the number of fillers.", "d Rel , d Arg , d P os indicate the TP-N2F decoder hyper-parameters, the dimension of relation vectors, the dimension of argument vectors, and the dimension of position vectors.", "In the experiment on the MathQA dataset, we use n F = 150, n R = 50, d F = 30, d R = 20, d Rel = 20, d Arg = 10, d P os = 5 and we train the model for 60 epochs with learning rate 0.00115.", "The reasoning module only contains one layer.", "As most of the math operators in this dataset are binary, we replace all operators taking three arguments with a set of binary operators based on hand-encoded rules, and for all operators taking one argument, a padding symbol is appended.", "For the baseline SEQ2PROG-orig, TP2LSTM and LSTM2TP, we use hidden size 100, single-direction, one-layer LSTM.", "For the SEQ2PROG-best, we performed a hyperparameter search on the hidden size for both encoder and decoder; the best score is reported.", "In the experiment on the AlgoLisp dataset, we use n F = 150, n R = 50, d F = 30, d R = 30, d Rel = 30, d Arg = 20, d P os = 5 and we train the model for 50 epochs with learning rate 0.00115.", "We also use one-layer in the reasoning module like in MathQA.", "For this dataset, most function calls take three arguments so we simply add padding symbols for those functions with fewer than three arguments." ]

[ 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.0714285671710968, 0.04999999329447746, 0.9767441749572754, 0.11320754140615463, 0.2083333283662796, 0.08510638028383255, 0.051282044500112534, 0.03999999538064003, 0.060606054961681366, 0.31578946113586426, 0.1304347813129425, 0.15686273574829102, 0.17142856121063232, 0, 0.06896550953388214, 0.13333332538604736, 0.1904761791229248, 0.12765957415103912, 0.1304347813129425, 0.3414634168148041, 0.10526315122842789, 0.1111111044883728, 0.10526315122842789, 0.10256409645080566, 0.1395348757505417, 0.29411762952804565, 0.13333332538604736, 0.07843136787414551, 0, 0.23529411852359772, 0.11764705181121826, 0.09302324801683426, 0.12765957415103912, 0.04444443807005882, 0.051282044500112534, 0.0555555522441864, 0.07692307233810425, 0, 0.04081632196903229, 0.10526315122842789, 0.09090908616781235, 0.04999999329447746, 0.1818181723356247, 0, 0.2222222238779068, 0.0555555522441864, 0.19512194395065308, 0.18518517911434174, 0, 0.1428571343421936 ]

[ "In this paper, we propose a new encoder-decoder model based on Tensor Product Representations for Natural- to Formal-language generation, called TP-N2F." ]

[ "This paper we present a defogger, a model that learns to predict future hidden information from partial observations.", "We formulate this model in the context of forward modeling and leverage spatial and sequential constraints and correlations via convolutional neural networks and long short-term memory networks, respectively.", "We evaluate our approach on a large dataset of human games of StarCraft: Brood War, a real-time strategy video game.", "Our models consistently beat strong rule-based baselines and qualitatively produce sensible future game states.", "We consider the problem of joint state estimation and next-state prediction in partially observable environments with complex dynamics.", "We take as a concrete example the problem of defogging in the real-time strategy (RTS) video game StarCraft, which we define as predicting the features of the game state that are hidden to the player.Forward modeling, the prediction of what is going to happen next, is a core enabler both for reactive control and for longer term planning.", "Many researchers are attempting to build and create algorithms that are able to model the future, especially in next frame video prediction and robotic planning BID9 BID0 One particular difficulty of forward modeling is to deal with the uncertainty of making a prediction with only a partial model and a partial view of the world.", "BID4 ; BID3 .In", "RTS games such as StarCraft, players must build an economy and control agents, called units, on a 2 dimensional grid to overcome their opponents. Several", "inherent limitations of any real-world setting are made explicit in such RTS games. First,", "by the \"fog of war\" which only allows players to see the surroundings of their own units and are thus unable to fully access the true game state. Second", ", the low-level dynamics are extremely complex, because several hundreds of agents interact together. However", ", there is an implicit spatio-temporal structure that makes long-term reasonning depend mostly on lower-resolution abstractions that can be obtained by averaging fine-grained characteristics over time and space. This poses", "a challenge for both human and computer players alike and predicting hidden information is a key step towards efficient planning in such environments, In this paper, as a first step towards learning a fully-featured forward model of the environment, the task we propose is to uncover hidden information and to predict the next state from observational data. We present", "a comprehensive analysis of a StarCraft Defogger, which predict features of the game at different levels of granularity: global features of the game such as the buildings of the opponent, and local features such as army density averaged by regions. Starting from", "a map of the environment subsampled in time and space, we propose a deep architecture of stacked long short-term memory cells applied convolutionally in an encoder-decoder architecture to predict the full state at different spatial resolutions. Individual layers", "of convolutional LSTMs encode the dynamics of the local features, and are aggregated in subsequent layers to model lower-resolution movements and global features. Trained on a large", "dataset of human replays BID8 , the model significantly outperforms strong rule-based baselines on several metrics." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.7777777910232544, 0.045454539358615875, 0.10810810327529907, 0.060606054961681366, 0, 0.1230769157409668, 0.16949151456356049, 0, 0.09090908616781235, 0, 0.045454539358615875, 0, 0.0833333283662796, 0.2153846174478531, 0.1666666567325592, 0.1538461446762085, 0.1395348757505417, 0.11428570747375488 ]

[ "This paper presents a defogger, a model that learns to predict future hidden information from partial observations, applied to a StarCraft dataset." ]

[ "In this work we study generalization of neural networks in gradient-based meta-learning by analyzing various properties of the objective landscapes.", "We experimentally demonstrate that as meta-training progresses, the meta-test solutions obtained by adapting the meta-train solution of the model to new tasks via few steps of gradient-based fine-tuning, become flatter, lower in loss, and further away from the meta-train solution.", "We also show that those meta-test solutions become flatter even as generalization starts to degrade, thus providing an experimental evidence against the correlation between generalization and flat minima in the paradigm of gradient-based meta-leaning.", "Furthermore, we provide empirical evidence that generalization to new tasks is correlated with the coherence between their adaptation trajectories in parameter space, measured by the average cosine similarity between task-specific trajectory directions, starting from a same meta-train solution.", "We also show that coherence of meta-test gradients, measured by the average inner product between the task-specific gradient vectors evaluated at meta-train solution, is also correlated with generalization.", "To address the problem of the few-shot learning, many meta-learning approaches have been proposed recently (Finn et al., 2017) , (Ravi and Larochelle, 2017) , (Rothfuss et al., 2018) , (Oreshkin et al., 2018) and (Snell et al., 2017) among others.", "In this work, we take steps towards understanding the characteristics of the landscapes of the loss functions, and their relation to generalization, in the context of gradient-based few-shot meta-learning.", "While we are interested in understanding the properties of optimization landscapes that are linked to generalization in gradient-based meta-learning in general, we focus our experimental work here within a setup that follows the recently proposed Model Agnostic Meta-Learning (MAML) algorithm (Finn et al., 2017) .", "The MAML algorithm is a good candidate for studying gradient-based meta-learning because of its independence from the underlying network architecture.", "Our main insights and contributions can be summarized as follows:", "1. As gradient-based meta-training progresses:", "• the adapted meta-test solutions become flatter on average, while the opposite occurs when using a finetuning baseline.", "• the adapted final solutions reach lower average support loss values, which never increases, while the opposite occurs when using a finetuning baseline.", "2. When generalization starts to degrade due to overtraining, meta-test solutions keep getting flatter, implying that, in the context of gradient-based meta-learning, flatness of minima is not correlated with generalization to new tasks.", "3. We empirically show that generalization to new tasks is correlated with the coherence between their adaptation trajectories, measured by the average cosine similarity between trajectory directions.", "Also correlated with generalization is the coherence between metatest gradients, measured by the average inner product between meta-test gradient vectors evaluated at meta-train solution.", "We also show that this metric is correlated to generalization for few-shot regression tasks where the model must learn to fit sine function curves.", "Furthermore, based on these observations, we take initial steps to propose a regularizer for MAML based training and provide experimental evidence for its effectiveness.", "We experimentally demonstrate that when using gradient-based meta-learning algorithms such as MAML, meta-test solutions, obtained after adapting neural networks to new tasks via few-shot learning, become flatter, lower in loss, and further away from the meta-train solution, as metatraining progresses.", "We also show that those meta-test solutions keep getting flatter even when generalization starts to degrade, thus providing an experimental argument against the correlation between generalization and flat minima.", "More importantly, we empirically show that generalization to new tasks is correlated with the coherence between their adaptation trajectories, measured by the average cosine similarity between the adaptation trajectory directions, but also correlated with the coherence between the meta-test gradients, measured by the average inner product between meta-test gradient vectors evaluated at meta-train solution.", "We also show this correlation for few-shot regression tasks.", "Based on these observations, we take first steps towards regularizing MAML based meta-training.", "As a future work, we plan to test the effectiveness of this regularizer on various datasets and meta-learning problem settings, architectures and gradient-based meta-learning algorithms.", "A ADDITIONAL EXPERIMENTAL DETAILS" ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.7222222089767456, 0.23529411852359772, 0.2448979616165161, 0.15094339847564697, 0.23255813121795654, 0.08888888359069824, 0.19512194395065308, 0.2142857164144516, 0.1621621549129486, 0, 0.09090908616781235, 0.05882352590560913, 0.05128204822540283, 0.21739129722118378, 0.1904761791229248, 0.1538461446762085, 0.14999999105930328, 0, 0.2142857164144516, 0.13333332538604736, 0.1090909019112587, 0.07692307233810425, 0, 0.19999998807907104, 0 ]

[ "We study generalization of neural networks in gradient-based meta- learning by analyzing various properties of the objective landscape." ]

[ "There have been multiple attempts with variational auto-encoders (VAE) to learn powerful global representations of complex data using a combination of latent stochastic variables and an autoregressive model over the dimensions of the data.", "However, for the most challenging natural image tasks the purely autoregressive model with stochastic variables still outperform the combined stochastic autoregressive models.", "In this paper, we present simple additions to the VAE framework that generalize to natural images by embedding spatial information in the stochastic layers.", "We significantly improve the state-of-the-art results on MNIST, OMNIGLOT, CIFAR10 and ImageNet when the feature map parameterization of the stochastic variables are combined with the autoregressive PixelCNN approach.", "Interestingly, we also observe close to state-of-the-art results without the autoregressive part.", "This opens the possibility for high quality image generation with only one forward-pass.\n", "In representation learning the goal is to learn a posterior latent distribution that explains the observed data well BID0 .", "Learning good representations from data can be used for various tasks such as generative modelling and semi-supervised learning (Kingma, 2013; BID14 BID14 BID23 .", "The decomposition of variational auto-encoders (VAE) (Kingma, 2013; BID14 provides the potential to disentangle the internal representation of the input data from local to global features through a hierarchy of stochastic latent variables.", "This makes the VAE an obvious candidate for learning good representations.", "However, in order to make inference tractable VAEs contain simplifying assumptions.", "This limits their ability to learn a good posterior latent representation.In complex data distributions with temporal dependencies (e.g. text, images and audio), the VAE assumption on conditional independence in the input distribution limits the ability to learn local structures.", "This has a significant impact on its generative performance, and thereby also the learned representations.", "Additionally, the one-layered VAE model with a N (0, I) latent prior poses serious constraints on the posterior complexity that the model is able to learn.", "A deep hierarchy of stochastic latent variables should endow the model with more expressiveness, but the VAE has a tendency to skip the learning of the higher representations since they pose a direct cost in its optimization term.There have been several attempts to eliminate the limitations of the VAE.", "Some concern formulating a more expressive variational distribution BID3 BID25 BID30 where other concerns learning a deeper hierarchy of latent variables .", "These contributions have resulted in better performance, but are still limited when modelling complex data distributions where a conditional independence does not apply.", "When parameterizing the VAE decoder with recurrent neural networks BID17 BID1 BID7 , the decoding architecture gets too powerful which results in unused latent stochastic variables .The", "limitations of the VAE have spawned interest towards other generative models such as Generative Adversarial Networks (GAN) BID8 and the autoregressive Pixel-CNN/PixelRNN models BID33 . These", "methods have proven powerful in learning good generative models, but the lack of stochastic latent variables makes them less suitable for representation learning purposes . Lately", ", we have seen several successful attempts to combine VAEs with PixelCNNs BID11 . This", "results Figure 1 : A visualization of FAME where the solid lines denote the variational approximation (inference/encoder/recognition) network and dashed lines denote the generative model (decoder) network for training. When", "performing reconstructions during training, the input image is concatenated with the output of the generative model (blue) and when generating the model follows a normal autoregressive sampling flow (red) while also using the stochastic latent variables z = z 1 , ..., z L . Both", "the variational approximation and the generative model follow a top-down hierarchical structure which enables precision weighted stochastic variables in the variational approximation.in a model where the global structure of the data is learned in the stochastic latent variables of the VAE and the local structure is learned in the PixelCNN. However", ", despite the additional complexity and potential extra expressiveness, these models do not outperform a simple autoregressive model BID32 .In this", "paper we present the Feature Map Variational Auto-Encoder (FAME) that combines the top-down variational approximation presented in the Ladder Variational Auto-Encoder (LVAE) ) with a spatial (feature map) representation of the stochastic latent variables and an autoregressive decoder. We show", "that (i) FAME", "outperforms previously state-of-the-art loglikelihood on MNIST, OMNIGLOT, CIFAR10 and ImageNet, (ii) FAME", "learns a deep hierarchy of stochastic latent variables without inactivated latent units, (iii) by", "removing the autoregressive decoder FAME performs close to previous state-of-the-art log-likelihood suggesting that it is possible to get good quality generation with just one forward pass.", "We have presented FAME, an extension to the VAE that significantly improve state-of-the-art performance on standard benchmark datasets.", "By introducing feature map representations in the latent stochastic variables in addition to top-down inference we have shown that the model is able to capture representations of complex image distributions while utilizing a powerful autoregressive architecture as a decoder.In order to analyze the contribution from the VAE as opposed to the autoregressive model, we have presented results without concatenating the input image when reconstructing and generating.", "This parameterization shows on par results with the previously state-of-the-art results without depending on the time consuming autoregressive generation.Further directions for FAME is to", "(i) test it on larger image datasets with images of a higher resolution,", "(ii) expand the model to capture other data modalities such as audio and text,", "(iii) combine the model in a semi-supervised framework." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.13636362552642822, 0.1249999925494194, 0.2222222238779068, 0.25641024112701416, 0.1538461446762085, 0, 0.1249999925494194, 0.1111111044883728, 0.0476190447807312, 0, 0, 0.16326530277729034, 0.27586206793785095, 0.21621620655059814, 0.07407406717538834, 0.05882352590560913, 0.05405404791235924, 0.04999999701976776, 0.10810810327529907, 0.052631575614213943, 0, 0.20512820780277252, 0.15686273574829102, 0.1860465109348297, 0.17142856121063232, 0.20408162474632263, 0.11764705926179886, 0.23076923191547394, 0.07407406717538834, 0.10256409645080566, 0.25, 0.1538461446762085, 0.1666666567325592, 0.2222222238779068, 0.1428571343421936, 0.1818181723356247 ]

[ "We present a generative model that proves state-of-the-art results on gray-scale and natural images." ]

[ "Plain recurrent networks greatly suffer from the vanishing gradient problem while Gated Neural Networks (GNNs) such as Long-short Term Memory (LSTM) and Gated Recurrent Unit (GRU) deliver promising results in many sequence learning tasks through sophisticated network designs.", "This paper shows how we can address this problem in a plain recurrent network by analyzing the gating mechanisms in GNNs.", "We propose a novel network called the Recurrent Identity Network (RIN) which allows a plain recurrent network to overcome the vanishing gradient problem while training very deep models without the use of gates.", "We compare this model with IRNNs and LSTMs on multiple sequence modeling benchmarks.", "The RINs demonstrate competitive performance and converge faster in all tasks.", "Notably, small RIN models produce 12%–67% higher accuracy on the Sequential and Permuted MNIST datasets and reach state-of-the-art performance on the bAbI question answering dataset.", "Numerous methods have been proposed for mitigating the vanishing gradient problem including the use of second-order optimization methods (e.g., Hessian-free optimization BID15 ), specific training schedules (e.g., Greedy Layer-wise training BID20 BID7 BID24 ), and special weight initialization methods when training on both plain FFNs and RNNs BID3 BID16 BID13 BID10 BID26 BID11 .Gated", "Neural Networks (GNNs) also help to mitigate this problem by introducing \"gates\" to control information flow through the network over layers or sequences. Notable", "examples include recurrent networks such as Long-short Term Memory (LSTM) BID8 , Gated Recurrent Unit (GRU) BID1 , and feedforward networks such as Highway Networks (HNs) BID21 , and Residual Networks (ResNets) BID5 . One can", "successfully train very deep models by employing these models, e.g., ResNets can be trained with over 1,000 layers. It has", "been demonstrated that removing (lesioning) or reordering (re-shuffling) random layers in deep feedforward GNNs does not noticeable affect the performance of the network BID23 Noticeably, one interpretation for this effect as given by BID4 is that the functional blocks in HNs or ResNets engage in an Unrolled Iterative Estimate (UIE) of representations and that layers in this block of HNs or ResNets iteratively refine a single set of representations.In this paper, we investigate if the view of Iterative Estimation (IE) can also be applied towards recurrent GNNs (Section 2.1). We present", "a formal analysis for GNNs by examining a dual gate design common in LSTM and GRU (Section 2.2). The analysis", "suggests that the use of gates in GNNs encourages the network to learn an identity mapping which can be beneficial in training deep architectures BID6 BID4 .We propose a", "new formulation of a plain RNN, called a Recurrent Identity Network (RIN) , that is encouraged to learn an identity mapping without the use of gates (Section 2). This network", "uses ReLU as the activation function and contains a set of non-trainable parameters. This simple", "yet effective method helps the plain recurrent network to overcome the vanishing gradient problem while it is still able to model long-range dependencies. This network", "is compared against two competing networks, the IRNN (Le et al., 2015) and LSTM, on several long sequence modeling tasks including the adding problem (Section 3.1), Sequential and Permuted MNIST classification tasks (Section 3.2), and bAbI question answering tasks (Section 3.3). RINs show faster", "convergence than IRNNs and LSTMs in the early stage of the training phase and reach competitive performance in all benchmarks. Note that the use", "of ReLU in RNNs usually leads to training instability, and therefore the network is sensitive to training hyperparameters. Our proposed RIN", "network demonstrates that a plain RNN does not suffer from this problem even with the use of ReLUs as shown in Section 3. We discuss further", "implications of this network and related work in Section 4.", "In this paper, we discussed the iterative representation refinement in RNNs and how this viewpoint could help in learning identity mapping.", "Under this observation, we demonstrated that the contribution of each recurrent step a GNN can be jointly determined by the representation that is formed up to the current step, and the openness of the carry gate in later recurrent updates.", "Note in Eq. 9, the element-wise multiplication of C t s selects the encoded representation that could arrive at the output of the layer.", "Thus, it is possible to embed a special function in C t s so that they are sensitive to certain pattern of interests.", "For example, in Phased LSTM, the time gate is inherently interested in temporal frequency selection BID17 .Motivated", "by the analysis presented in Section 2, we propose a novel plain recurrent network variant, the Recurrent Identity Network (RIN), that can model long-range dependencies without the use of gates. Compared", "to the conventional formulation of plain RNNs, the formulation of RINs only adds a set of non-trainable weights to represent a \"surrogate memory\" component so that the learned representation can be maintained across two recurrent steps.Experimental results in Section 3 show that RINs are competitive against other network models such as IRNNs and LSTMs. Particularly", ", small RINs produce 12%-67% higher accuracy in the Sequential and Permuted MNIST. Furthermore", ", RINs demonstrated much faster convergence speed in early phase of training, which is a desirable advantage for platforms with limited computing resources. RINs work", "well without advanced methods of weight initializations and are relatively insensitive to hyperparameters such as learning rate, batch size, and selection of optimizer. This property", "can be very helpful when the time available for choosing hyperparameters is limited. Note that we", "do not claim that RINs outperform LSTMs in general because LSTMs may achieve comparable performance with finely-tuned hyperparameters.The use of ReLU in RNNs might be counterintuitive at first sight because the repeated application of this activation is more likely causing gradient explosion than conventional choices of activation function, such as hyperbolic tangent (tanh) function or sigmoid function. Although the", "proposed IRNN BID13 reduces the problem by the identity initialization, in our experiments, we usually found that IRNN is more sensitive to training parameters and more unstable than RINs and LSTMs. On the contrary", ", feedforward models that use ReLU usually produce better results and converge faster than FFNs that use the tanh or sigmoid activation function. In this paper,", "we provide a promising method of using ReLU in RNNs so that the network is less sensitive to the training conditions. The experimental", "results also support the argument that the use of ReLU significantly speeds up the convergence.During the development of this paper, a recent independent work BID27 presented a similar network formulation with a focus on training of deep plain FFNs without skip connections. DiracNet uses the", "idea of ResNets where it assumes that the identity initialization can replace the role of the skip-connection in ResNets. DiracNet employed", "a particular kind of activation function -negative concatenated ReLU (NCReLU), and this activation function allows the layer output to approximate the layer input when the expectation of the weights are close to zero. In this paper, we", "showed that an RNN can be trained without the use of gates or special activation functions, which complements the findings and provides theoretical basis in BID27 .We hope to see more", "empirical and theoretical insights that explains the effectiveness of the RIN by simply embedding a non-trainable identity matrix. In future, we will", "investigate the reasons for the faster convergence speed of the RIN during training. Furthermore, we will", "investigate why RIN can be trained stably with the repeated application of ReLU and why it is less sensitive to training parameters than the two other models.A ALGEBRA OF EQS. 8-9Popular GNNs such", "as LSTM, GRU; and recent variants such as the Phased-LSTM BID17 , and Intersection RNN BID2 , share the same dual gate design described as follows: DISPLAYFORM0 where t ∈ [1, T ], H t = σ(x t , h t−1 ) represents the hidden transformation, T t = τ (x t , h t−1 ) is the transform gate, and C t = φ(x t , h t−1 ) is the carry gate. σ, τ and φ are recurrent", "layers that have their trainable parameters and activation functions. represents element-wise", "product operator. Note that h t may not be", "the output activation at the recurrent step t. For example in LSTM, h t", "represents the memory cell state. Typically, the elements", "of transform gate T t,k and carry gate C t,k are between 0 (close) and 1 (open), the value indicates the openness of the gate at the kth neuron. Hence, a plain recurrent", "network is a subcase of Eq. 14 when T t = 1 and C t = 0.Note that conventionally, the initial hidden activation h 0 is 0 to represent a \"void state\" at the start of computation. For h 0 to fit into Eq.", "4's framework, we define an auxiliary state h −1 as the previous state of h 0 , and T 0 = 1, C 0 = 0. We also define another", "auxiliary state h T +1 = h T , T T +1 = 0, and C T +1 = 1 as the succeeding state of h T .Based on the recursive", "definition in Eq. 4, we can write the final layer output h T as follows: DISPLAYFORM1 where we use to represent element-wise multiplication over a series of terms.According to Eq. 3, and supposing that Eq. 5 fulfills the Eq. 1, we can use a zero-mean residual t for describing the difference between the outputs of recurrent steps: DISPLAYFORM2 Then we can rewrite Eq. 16 as: DISPLAYFORM3 Substituting Eq. 18 into Eq. 15: DISPLAYFORM4 We can rearrange Eqn. 20 to DISPLAYFORM5 The", "term λ in Eq. 23 can be reorganized to, DISPLAYFORM6 B DETAILS IN THE ADDING PROBLEM EXPERIMENTS Average Estimation Error RIN 2-100 1st IRNN 2-100 1st LSTM 2-100 1st 0 100 200 300 400 500 600 700 800 layer 2 step index RIN 2-100 2nd IRNN 2-100 2nd LSTM 2-100 2nd DISPLAYFORM7 DISPLAYFORM8" ]

[ 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.24242423474788666, 0.2448979616165161, 1, 0.0476190447807312, 0, 0.07843136787414551, 0.21052631735801697, 0.1538461446762085, 0.07017543166875839, 0.11764705181121826, 0.14000000059604645, 0.0416666604578495, 0.4285714328289032, 0.4912280738353729, 0.13636362552642822, 0.3921568691730499, 0.05882352590560913, 0.16326530277729034, 0.2083333283662796, 0.290909081697464, 0.10256409645080566, 0.0416666604578495, 0.16129031777381897, 0.08163265138864517, 0.11764705181121826, 0.04444443807005882, 0.48275861144065857, 0.21052631735801697, 0.04651162400841713, 0.11320754140615463, 0.11538460850715637, 0.08888888359069824, 0.09876542538404465, 0.13793103396892548, 0.11320754140615463, 0.23529411852359772, 0.2686567008495331, 0.08695651590824127, 0.17543859779834747, 0.2711864411830902, 0.11999999731779099, 0.1395348757505417, 0.16393442451953888, 0.05128204822540283, 0, 0, 0.09756097197532654, 0.0555555522441864, 0.18518517911434174, 0.16129031777381897, 0.11538460850715637, 0.08510638028383255, 0.1538461446762085, 0 ]

[ "We propose a novel network called the Recurrent Identity Network (RIN) which allows a plain recurrent network to overcome the vanishing gradient problem while training very deep models without the use of gates." ]

[ "Recently, there has been a surge in interest in safe and robust techniques within reinforcement learning (RL). \n", "Current notions of risk in RL fail to capture the potential for systemic failures such as abrupt stoppages from system failures or surpassing of safety thresholds and the appropriate responsive controls in such instances.", "We propose a novel approach to fault-tolerance within RL in which the controller learns a policy can cope with adversarial attacks and random stoppages that lead to failures of the system subcomponents.", "The results of the paper also cover fault-tolerant (FT) control so that the controller learns to avoid states that carry risk of system failures.", "By demonstrating that the class of problems is represented by a variant of SGs, we prove the existence of a solution which is a unique fixed point equilibrium of the game and characterise the optimal controller behaviour.", "We then introduce a value function approximation algorithm that converges to the solution through simulation in unknown environments.", "Reinforcement learning (RL) provides the promise of adaptive agents being able to discover solutions merely through repeated interaction with their environment.", "RL has been deployed in a number of real-world settings in which, using RL, an adaptive agent learns to perform complex tasks, often in environments shared by human beings.", "Large scale factory industrial applications, traffic light control (Arel et al., 2010) , robotics (Deisenroth et al., 2013) and autonomous vehicles (Shalev-Shwartz et al., 2016) are notable examples of settings to which RL methods have been applied.", "Numerous automated systems are however, susceptible to failures and unanticipated outcomes.", "Moreover, many real-world systems amenable to RL suffer the potential for random stoppages and abrupt failures; actuator faults, failing mechanical system components, sensor failures are few such examples.", "In these settings, executing preprogrammed behaviours or policies that have been trained in idealised simulated environments can prove vastly inadequate for the task of ensuring the safe execution of tasks.", "Consequently, in the presence of such occurrences, the deployment of RL agents introduces a risk of catastrophic outcomes whenever the agent is required to act so as to avoid adverse outcomes in unseen conditions.", "The important question of how to control the system in a way that is both robust against systemic faults and, minimises the risk of faults or damage therefore arises.", "In response to the need to produce RL algorithms that execute tasks with safety guarantees, a significant amount of focus has recently been placed on safe execution, robust control and riskminimisation (Garcıa and Fernández, 2015) .", "Examples include H ∞ control (Morimoto and Doya, 2001) , coherent risk, conditional value at risk (Tamar et al., 2015) .", "In general, these methods introduce an objective 1 defined with an expectation measure that either penalises actions that lead to greater uncertainty or embeds a more pessimistic view of the world (for example, by biasing the transition predictions towards less desirable states).", "In both cases, the resulting policies act more cautiously over the horizon of the problem as compared to policies trained with a standard objective function.", "Despite the recent focus on safe methods within RL, the question of how to train an RL agent that can cope with random failures remains unaddressed.", "In particular, at present the question of how to produce an RL policy that can cope with an abrupt failure of some system subcomponent has received no systematic treatment.", "Similarly, the task of addressing how to produce RL policies that account for the risk of states in which such failures occur has not been addressed.", "In this paper, we for the first time produce a method that learns optimal policies in response to random and adversarial systems attacks that lead to stoppages of system (sub)components that may produce adverse events.", "Our method works by introducing an adversary that seeks to determine a stopping criterion to stop the system at states that lead to the worst possible (overall) outcomes for the controller.", "Using a game-theoretic construction, we then show how a policy that is robust against adversarial attacks that lead to abrupt failure can be learned by an adaptive agent using an RL updating method.", "In particular, the introduction of an adversary that performs attacks at states that lead to worst outcomes generates experiences for the adaptive RL agent to learn a best-response policy against such scenarios.", "To tackle this problem, we construct a novel two-player stochastic game (SG) in which one of the players, the controller, is delegated the task of learning to modify the system dynamics through its actions that maximise its payoff and an adversary or 'stopper' that enacts a strategy that stops the system in such a way that maximises the controller's costs.", "This produces a framework that finds optimal policies that are robust against stoppages at times that pose the greatest risk of catastrophe.", "The main contribution of the paper is to perform the first systematic treatment of the problem of robust control under worst-case failures.", "In particular, we perform a formal analysis of the game between the controller and the stopper.", "Our main results are centered around a minimax proof that establishes the existence of a value of the game.", "This is necessary for simulating the stopping action to induce fault-tolerance.", "Although minimax proofs are well-known in game theory (Shapley, 1953; Maitra and Parthasarathy, 1970; Filar et al., 1991) , replacing a player's action set with stopping rules necessitates a minimax proof (which now relies on a construction of open sets) which markedly differs to the standard methods within game theory.", "Additionally, crucial to our analysis is the characterisation of the adversary optimal stopping rule (Theorem 3).", "Our results tackle optimal stopping problems (OSPs) under worst-case transitions.", "OSPs are a subclass of optimal stochastic control (OSC) problems in which the goal is to determine a criterion for stopping at a time that maximises some state-dependent payoff (Peskir and Shiryaev, 2006) .", "The framework is developed through a series of theoretical results: first, we establish the existence of a value of the game which characterises the payoff for the saddle point equilibrium (SPE).", "Second, we prove a contraction mapping property of a Bellman operator of the game and that the value is a unique fixed point of the operator.", "Third, we prove the existence and characterise the optimal stopping time.", "We then prove an equivalence between the game of control and stopping and worst-case OSPs and show that the fixed point solution of the game solves the OSP.", "Finally, using an approximate dynamic programming method, we develop a simulation-based iterative scheme that computes the optimal controls.", "The method applies in settings in which neither the system dynamics nor the reward function are known.", "Hence, the agent need only observe its realised rewards by interacting with the environment.", "In this paper, we tackled the problem of fault-tolerance within RL in which the controller seeks to obtain a control that is robust against catastrophic failures.", "To formally characterise the optimal behaviour, we constructed a new discrete-time SG of control and stopping.", "We established the existence of an equilibrium value then, using a contraction mapping argument, showed that the game can be solved by iterative application of a Bellman operator and constructed an approximate dynamic programming algorithm so that the game can be solved by simulation.", "Assumption A.2.", "Ergodicity: i) Any invariant random variable of the state process is P −almost surely (P −a.s.) a constant.", "Assumption A.3.", "Markovian transition dynamics: the transition probability function P satisfies the following equality:", "Assumption A.4.", "The constituent functions {R, G} in J are square integrable: that is, R, G ∈ L 2 (µ)." ]

[ 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.07692307233810425, 0.10526315122842789, 0.2631579041481018, 0.13333332538604736, 0.054054051637649536, 0, 0, 0, 0.0476190447807312, 0.09999999403953552, 0.1621621549129486, 0, 0, 0.05714285373687744, 0.0476190447807312, 0.06896550953388214, 0, 0, 0.05882352590560913, 0, 0, 0.20000000298023224, 0, 0.05128204822540283, 0, 0.036363635212183, 0.06896550953388214, 0.2222222238779068, 0.08695651590824127, 0, 0.09999999403953552, 0.03703703358769417, 0, 0.10526315122842789, 0.04999999701976776, 0.05882352590560913, 0.0714285671710968, 0.10526315122842789, 0.06666666269302368, 0, 0.0833333283662796, 0, 0.05882352590560913, 0.07999999821186066, 0.0476190447807312, 0.06896550953388214, 0, 0.07407406717538834 ]

[ "The paper tackles fault-tolerance under random and adversarial stoppages." ]

[ "We propose a novel framework to generate clean video frames from a single motion-blurred image.\n", "While a broad range of literature focuses on recovering a single image from a blurred image, in this work, we tackle a more challenging task i.e. video restoration from a blurred image.", "We formulate video restoration from a single blurred image as an inverse problem by setting clean image sequence and their respective motion as latent factors, and the blurred image as an observation.", "Our framework is based on an encoder-decoder structure with spatial transformer network modules to restore a video sequence and its underlying motion in an end-to-end manner.", "We design a loss function and regularizers with complementary properties to stabilize the training and analyze variant models of the proposed network.", "The effectiveness and transferability of our network are highlighted through a large set of experiments on two different types of datasets: camera rotation blurs generated from panorama scenes and dynamic motion blurs in high speed videos.", "Our code and models will be publicly available.", "Capturing an image is not an instant process; to capture enough photons, the photosensitive elements of a camera have to be exposed to light for a certain interval of time, called exposure time.", "Therefore, during this interval if an object is moving in the observed scene or the camera is undergoing an arbitrary motion, the resulting image will contain a blurring artifact known as motion blur.", "In general, motion blur is an unwanted behaviour in vision applications e.g.image editing (Gunturk & Li, 2012 ), visual SLAM (Lee et al., 2011 and 3D reconstruction (Seok Lee & Mu Lee, 2013) , as it degrades the visual quality of images.", "To cope with this type of artifact, image deblurring aims to restore a sharp image from a blurred image.", "This problem is known to be ill-posed since the blur kernel used for deconvolution is generally assumed to be unknown.", "Earlier studies assume a uniform-blur over the image to simplify the estimation of the single deconvolution blur kernel used to remove the blur (Fergus et al., 2006; Cho & Lee, 2009; Levin et al., 2009) .", "Even though the methods deploy deblurring tasks with uniform-blur assumption, the assumption is often violated in practice.", "For instance, when the blur is caused by out-of-plane camera rotation, the blur pattern becomes spatially variant.", "Moreover, the problem is more complex when objects in a scene are moving i.e.dynamic blur.", "While previous literature focuses on recovering a sharp image from a blurred image, we tackle a more challenging task i.e.video restoration from a blurred image.", "Restoring the underlying image sequence of a blurred image requires both contents and motion prediction.", "We formulate video restoration from a blurred image as an inverse problem where a clean sequence of images and their motion as latent factors, and a blurred image as an observation.", "Some of previous deblurring approaches (Hyun Kim & Mu Lee, 2014; Zhang & Yang, 2015; Sellent et al., 2016; Ren et al., 2017; Park & Mu Lee, 2017 ) also estimate the underlying motion in a blurred image, however, their goal remains in single frame restoration.", "Recently Jin et al. (2018) proposed to extract video frames from a single motion-blurred image.", "Their approach is close to image translation model without inferring underlying motions between the latent frames.", "Purohit et al. (2019) addressed this issue by estimating pixel level motion from a given blurred input.", "However, their model is still prone to sequential error propagation as frames are predicted in a sequential manner using a deblurred middle frame.", "In this paper, we propose a novel framework to generate a clean sequence of images from a single motion-blurred image.", "Our framework is based on a single encoder-decoder structure with Spatial Transformer Network modules (STN) and Local Warping layers (LW) to restore an image sequence and its underlying motion.", "Specifically, a single encoder is used to extract intermediate features which are passed to multiple decoders with predicted motion from STN and LW modules to generate a sequence of deblurred images.", "We evaluate our model on two types of motion blur.", "For rotation blur, which is caused by abrupt camera motion, we generated a synthetic dataset from panoramic images (J. Xiao & Torralba., 2012) .", "For dynamic blur caused by fast moving objects in a scene, we used a high speed video dataset (Nah et al., 2017) .", "The proposed model is evaluated on the panorama and the high speed video datasets under various motion patterns.", "Both the quantitative metrics and qualitative results highlight that our method is more robust and performs favorably against the competing approach (Jin et al., 2018) 1 .", "For further investigation, we demonstrate the transferability of our model by cross-dataset evaluation.", "We also propose a simpler and lighter variation of our model guiding that our approach is flexible and can be easily extended to arbitrary number of frame prediction model with negligible performance trade-off.", "In short, our contributions are as follows.", "1) We propose a novel unified architecture to restore clean video frames from a single motion-blurred image in an end-to-end manner.", "2) Loss terms are designed to stably train the proposed network.", "3) We perform thorough experiments to analyze the transferability and flexibility of the proposed architecture.", "4) The performance of our model quantitatively and qualitatively performs favorably against the competing approach.", "Moreover due to flexibility of our model, we show that our approach is robust to heavy blurs where the previous approach fails." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0 ]

[ 0.5454545617103577, 0.27272728085517883, 0.3255814015865326, 0.2790697515010834, 0.10526315122842789, 0.11999999731779099, 0, 0.1304347813129425, 0.1702127605676651, 0.09999999403953552, 0.1764705777168274, 0, 0.12765957415103912, 0.05882352590560913, 0, 0.11428570747375488, 0.20512819290161133, 0.1249999925494194, 0.2926829159259796, 0.10344827175140381, 0.42424240708351135, 0.11764705181121826, 0.11428570747375488, 0.20512819290161133, 0.3333333134651184, 0.17391303181648254, 0.1304347813129425, 0.0714285671710968, 0.0952380895614624, 0.14999999105930328, 0.05714285373687744, 0.04651162400841713, 0, 0.12765957415103912, 0, 0.7894736528396606, 0, 0.1249999925494194, 0, 0.05405404791235924 ]

[ "We present a novel unified architecture that restores video frames from a single motion-blurred image in an end-to-end manner." ]

[ "High performance of deep learning models typically comes at cost of considerable model size and computation time.", "These factors limit applicability for deployment on memory and battery constraint devices such as mobile phones or embedded systems.", "In this work we propose a novel pruning technique that eliminates entire filters and neurons according to their relative L1-norm as compared to the rest of the network, yielding more compression and decreased redundancy in the parameters.", "The resulting network is non-sparse, however, much more compact and requires no special infrastructure for its deployment.", "We prove the viability of our method by achieving 97.4%, 47.8% and 53% compression of LeNet-5, ResNet-56 and ResNet-110 respectively, exceeding state-of-the-art compression results reported on ResNet without losing any performance compared to the baseline.", "Our approach does not only exhibit good performance, but is also easy to implement on many architectures.", "While deep learning models have become the method of choice for a multitude of applications, their training requires a large number of parameters and extensive computational costs (energy, memory footprint, inference time).", "This limits their deployment on storage and battery constraint devices, such as mobile phones and embedded systems.", "To compress deep learning models without loss in accuracy, previous work proposed pruning weights by optimizing network's complexity using second order derivative information BID1 BID4 .", "While second order derivative introduces a high computational overhead, BID7 BID9 explored low rank approximations to reduce the size of the weight tensors.Another line of work BID3 BID14 , proposed to prune individual layer weights with the lowest absolute value (nonstructural sparsification of layer weights).", "BID2 followed the same strategy while incorporating quantization and Huffman coding to further boost compression.", "While the aforementioned methods considered every layer independently, BID12 proposed to prune the network weights in a class-blind manner, e.g. individual layer weights are pruned according to their magnitude as compared to all weights in the network.Noteworthy, all approaches that prune weights non-structurally, generally result in high sparsity models that require dedicated hardware and software.", "Structured pruning alleviates this by removing whole filters or neurons, producing a non-sparse compressed model.", "In this regard, BID11 proposed channel-wise pruning according to the L1-norm of the corresponding filter.", "BID15 learned a compact model based on learning structured sparsity of different parameters.", "A data-free algorithm was implemented to remove redundant neurons iteratively on fully connected layers in BID13 .", "In BID6 , connections leading to weak activations were pruned.", "Finally, BID16 pruned neurons by measuring their importance with respect to the penultimate layer.", "Generally, in structured pruning, each layer is pruned separately, which requires calculation of layer importance before training.", "This work features two key components:", "a) Blindness: all layers are considered simultaneously; blind pruning was first introduced by BID12 to prune individual weights;", "b) Structured Pruning: removal of entire filters instead of individual weights.", "To the best of our knowledge, we are the first to use these two components together to prune filters based on their relative L1-norm compared to the sum of all filters' L1-norms across the network, instead of pruning filters according to their L1-norm within the layer BID11 , inducing a global importance score for each filter.", "The contribution of this paper is two-fold:", "i) Proposing a structured class-blind pruning technique to compress the network by removing whole filters and neurons, which results in a compact non-sparse network with the same baseline performance.", "ii) Introducing a visualization of global filter importance to devise the pruning percentage of each layer.As a result, the proposed approach achieves higher compression gains with higher accuracy compared to the state-of-the-art results reported on ResNet-56 and ResNet-110 on the CIFAR10 dataset BID8 .", "We presented a novel structured pruning method to compress neural networks without losing accuracy.", "By pruning layers simultaneously instead of looking at each layer individually, our method combines all filters and output features of all layers and prunes them according to a global threshold.", "We have surpassed state-of-the-art compression results reported on ResNet-56 and ResNet-110 on CIFAR-10 BID16 , compressing more than 47% and 53% respectively.", "Also, we showed that only 11K parameters are sufficient to exceed the baseline performance on LeNet-5, compressing more than 97%.", "To realize the advantages of our method, no customized hardware or libraries are needed.", "It is worth to say that due to removing whole filters and neurons, the pruning percentage reflects the effective model compression percentage.", "For the future work, we are dedicated to proving the applicability of our method on several different architectures and datasets.", "Hence, we plan to experiment on VGG-16, ResNet on ImageNet and/or other comparable architectures." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0 ]

[ 0, 0, 0.25531914830207825, 0, 0.08510638028383255, 0.06451612710952759, 0.04651162400841713, 0, 0.05128204822540283, 0.07407406717538834, 0.06896550953388214, 0.10526315122842789, 0.20689654350280762, 0.1428571343421936, 0.14814814925193787, 0.06666666269302368, 0.0833333283662796, 0.0714285671710968, 0.06666666269302368, 0, 0.1249999925494194, 0, 0.10344827175140381, 0, 0.29999998211860657, 0.11999999731779099, 0.5714285373687744, 0.14999999105930328, 0.05882352590560913, 0.05882352590560913, 0, 0.12121211737394333, 0.060606054961681366, 0.07407406717538834 ]

[ "We propose a novel structured class-blind pruning technique to produce highly compressed neural networks." ]

[ "The recent success of neural networks for solving difficult decision tasks has incentivized incorporating smart decision making \"at the edge.\"", "However, this work has traditionally focused on neural network inference, rather than training, due to memory and compute limitations, especially in emerging non-volatile memory systems, where writes are energetically costly and reduce lifespan.", "Yet, the ability to train at the edge is becoming increasingly important as it enables applications such as real-time adaptability to device drift and environmental variation, user customization, and federated learning across devices.", "In this work, we address four key challenges for training on edge devices with non-volatile memory: low weight update density, weight quantization, low auxiliary memory, and online learning.", "We present a low-rank training scheme that addresses these four challenges while maintaining computational efficiency.", "We then demonstrate the technique on a representative convolutional neural network across several adaptation problems, where it out-performs standard SGD both in accuracy and in number of weight updates.", "Deep neural networks have shown remarkable performance on a variety of challenging inference tasks.", "As the energy efficiency of deep-learning inference accelerators improves, some models are now being deployed directly to edge devices to take advantage of increased privacy, reduced network bandwidth, and lower inference latency.", "Despite edge deployment, training happens predominately in the cloud.", "This limits the privacy advantages of running models on-device and results in static models that do not adapt to evolving data distributions in the field.", "Efforts aimed at on-device training address some of these challenges.", "Federated learning aims to keep data on-device by training models in a distributed fashion (Konecný et al., 2016) .", "On-device model customization has been achieved by techniques such as weight-imprinting (Qi et al., 2018) , or by retraining limited sets of layers.", "On-chip training has also been demonstrated for handling hardware imperfections (Zhang et al., 2017; Gonugondla et al., 2018) .", "Despite this progress with small models, on-chip training of larger models is bottlenecked by the limited memory size and compute horsepower of edge processors.", "Emerging non-volatile (NVM) memories such as resistive random access memory (RRAM) have shown great promise for energy and area-efficient inference (Yu, 2018) .", "However, on-chip training requires a large number of writes to the memory, and RRAM writes cost significantly more energy than reads (e.g., 10.9 pJ/bit versus 1.76 pJ/bit (Wu et al., 2019) ).", "Additionally, RRAM endurance is on the order of 10 6 writes (Grossi et al., 2019) , shortening the lifetime of a device due to memory writes for on-chip training.", "In this paper, we present an online training scheme amenable to NVM memories to enable next generation edge devices.", "Our contributions are (1) an algorithm called Streaming Kronecker Sum Approximation (SKS), and its analysis, which addresses the two key challenges of low write density and low auxiliary memory; (2) two techniques \"gradient max-norm\" and \"streaming batch norm\" to help training specifically in the online setting; (3) a suite of adaptation experiments to demonstrate the advantages of our approach.", "We demonstrated the potential for SKS to solve the major challenges facing online training on NVM-based edge devices: low write density and low auxiliary memory.", "SKS is a computationallyefficient, memory-light algorithm capable of decoupling batch size from auxiliary memory, allowing larger effective batch sizes, and consequently lower write densities.", "Additionally, we noted that SKS may allow for training under severe weight quantization constraints as rudimentary gradient accumulations are handled by the L, R matrices, which can have high bitwidths (as opposed to SGD, which may squash small gradients to 0).", "We found expressions for when SKS might have better convergence properties.", "Across a variety of online adaptation problems and a large-scale transfer learning demonstration, SKS was shown to match or exceed the performance of SGD while using a small fraction of the number of updates.", "Finally, we suspect that these techniques could be applied to a broader range of problems.", "Auxiliary memory minimization may be analogous to communication minimization in training strategies such as federated learning, where gradient compression is important." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0 ]

[ 0.1904761791229248, 0.22641508281230927, 0.11764705181121826, 0.2916666567325592, 0.2702702581882477, 0.1599999964237213, 0.1666666567325592, 0.11764705181121826, 0.19354838132858276, 0.13636362552642822, 0.1875, 0.1463414579629898, 0, 0.10256409645080566, 0.13333332538604736, 0.09090908616781235, 0.0714285671710968, 0.2083333283662796, 0.19999998807907104, 0.138888880610466, 0.35555556416511536, 0, 0.13333332538604736, 0.12121211737394333, 0.03999999538064003, 0.10810810327529907, 0.1904761791229248 ]

[ "We use Kronecker sum approximations for low-rank training to address challenges in training neural networks on edge devices that utilize emerging memory technologies." ]

[ "Knowledge extraction techniques are used to convert neural networks into symbolic descriptions with the objective of producing more comprehensible learning models.", "The central challenge is to find an explanation which is more comprehensible than the original model while still representing that model faithfully.", "The distributed nature of deep networks has led many to believe that the hidden features of a neural network cannot be explained by logical descriptions simple enough to be understood by humans, and that decompositional knowledge extraction should be abandoned in favour of other methods.", "In this paper we examine this question systematically by proposing a knowledge extraction method using \\textit{M-of-N} rules which allows us to map the complexity/accuracy landscape of rules describing hidden features in a Convolutional Neural Network (CNN).", "Experiments reported in this paper show that the shape of this landscape reveals an optimal trade off between comprehensibility and accuracy, showing that each latent variable has an optimal \\textit{M-of-N} rule to describe its behaviour.", "We find that the rules with optimal tradeoff in the first and final layer have a high degree of explainability whereas the rules with the optimal tradeoff in the second and third layer are less explainable.", "The results shed light on the feasibility of rule extraction from deep networks, and point to the value of decompositional knowledge extraction as a method of explainability.", "Recently there has been an increase in interest in explainable Artificial Intelligence (AI).", "Although in the past decade there have been major advances in the performance of neural network models, these models tend not to be explainable (Andrew Gordon Wilson, 2017) .", "In large part, this is due to the use of very large networks, specifically deep networks, which rely on distributed representations to model data accurately BID11 .", "In contrast with symbolic AI, in which specific features are often hand picked for a problem, or symbolic Machine Learning (ML), which takes a localist approach BID15 , the features used by a distributed representation do not necessarily correlate with obviously identifiable features of the data.", "A distributed representation may owe its strength to weak statistical correlations that a human would not be able to detect or describe in any comprehensible way.Knowledge extraction seeks to increase the explainability of neural networks by attempting to uncover the knowledge that a neural network has learned implicitly in its weights.", "One way of doing this is to translate trained neural networks into a set of symbolic rules or decision trees similar to the ones found in symbolic AI, ML and logic programming BID16 BID7 .", "Rule extraction techniques have been around for decades BID20 ) with a number of rule extraction algorithms having been developed over the years BID12 BID4 BID22 ) (d'Avila BID5 .", "These techniques generally take one of two approaches: decompositional, in which the parameters of the network are used to generate rules, or pedagogical, in which the behaviour of the network is used to generate rules BID1 .", "In either case, the major issue with rule extraction is the complexity of the extracted rules.", "Even if it is possible to find a symbolic system which describes exactly a neural network (for example, feedforward, Boolean, deterministic networks can always be written as a logic program), a very large rule set derived from a very large CNN may be no more comprehensible than the original network.Perhaps the main reason knowledge extraction proves difficult (and in particular decompositional methods of extraction) is the distributed representations found in neural networks BID11 .", "This means that important concepts which can be used for reasoning are not always represented by single neurons but by patterns of activity over many neurons.", "It has been argued that the distributed nature of neural networks plays an important part in many of their capabilities BID19 .", "Distributed representations have been identified as one of the fundamental properties of connectionism BID18 .", "This has led many to conclude that attempting to explain latent features using symbolic knowledge extraction is a dead end, and that methods akin to distillation should be adopted instead BID7 .", "Distillation has also been proposed as a method for improving robustness but it's efficacy has been questioned BID13 BID3 .", "Other approaches take a more practical view.", "Rather than attempting to open the black box, one may settle for some guarantees on the network's behaviour, or for visualizations seeking to explain individual classifications rather than the learned model BID9 BID17 BID10 In this paper, we develop a method for empirically examining the explainability of the latent variables in neural networks.", "We use rule extraction by searching through a space of M-of-N rules BID20 ) describing a latent variable, and measuring the error and complexity of each rule.", "By selecting various error/complexity trade-offs, we are able to map out a rule extraction landscape which shows the relationship between how complex the extracted rules are allowed to be and how accurately they capture the behaviour of a network.", "When applied to a standard 4-layer CNN trained on fashion MNIST, we find that some layers have very accurate rules whereas this is not the case for others even when using very complex rules.", "The discovery of a 'critical point' on the rule extraction landscape shows that there is an ideal M-of-N rule to describe each latent variable.", "The accuracy of those rules depends highly on the variable that we are attempting to describe, with the overall explainability trends differing greatly between layers and architectures.", "All layers showed similarly shaped curves but in the convolutional layers the rules extracted with no penalty in complexity were much more complex relatively than the ones extracted from the fully connected layers with relative complexities over 0.4 in the convolutional layers and complexities of under 0.2 in the fully connected layers.", "Additionally, it was possible to find rules with near 0% error in the first and final layer whereas rules from the second and third layer could not do much better than 15% error.In Section 2 we give a brief overview of previous algorithms used for knowledge extraction.", "In Section 3 we give definitions of accuracy and complexity for M-of-N rules and outline the extraction process.", "In Section 4 we give the experimental results of our rule extraction process for the mapping of the accuracy/complexity landscape before concluding in Section 5.", "The black box problem of neural networks presents an obstacle to their deployment into society.", "The black box problem has been an issue for neural networks since their creation, but as neural networks have become more integrated into society, the need for explainability has attracted considerably more attention.", "The success of knowledge extraction in this endeavor has overall been mixed with most large neural networks today remaining difficult to interpret and explain.", "Traditionally knowledge extraction has been a commonly used paradigm and it has been applied to various tasks.", "Critics, however, point out that the distributed nature of neural networks makes the specific method of decomposition rule extraction unfeasible as individual latent features and unlikely to represent anything of significance.", "We test this claim by applying a novel search method for M-of-N rules to explain the latent features of a CNN, and find that generally latent features can be described by an 'optimal' rule representing an ideal error/complexity trade-off for the explanation.", "We do this by including rule complexity as an explicit measure in the search for extracted rules.", "The large discrepancy in this trade-off between neurons in different layers, neurons in different layers with different architectures, and even different neurons in the same layer, suggests that rule extraction as a general technique is unlikely to provide adequate descriptions for all, or even most latent variables.However, the fact that in many cases the explanations can be made much simpler without reducing the accuracy of the rules suggests that rule extraction can be a useful tool when examining networks with features that are likely to be easily understandable.", "These results indicate that decompositional rule extraction may still be an important tool for understanding the behaviour of networks.", "Further research would examine the effects on the accuracy/interpretability landscape of using different transfer functions, other data sets, different architectures, and various forms of regularization of the learning." ]

[ 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.10256409645080566, 0.052631575614213943, 0.3214285671710968, 0.31372547149658203, 0.12244897335767746, 0.23255813121795654, 0.19512194395065308, 0.06666666269302368, 0.1818181723356247, 0.1463414579629898, 0.178571417927742, 0.16393442451953888, 0.20408162474632263, 0.13636362552642822, 0.23255813121795654, 0.1875, 0.1538461446762085, 0.0952380895614624, 0.15789473056793213, 0.12903225421905518, 0.1304347813129425, 0.05714285373687744, 0.07999999821186066, 0.1904761791229248, 0.19512194395065308, 0.2745097875595093, 0.1599999964237213, 0.1463414579629898, 0.1818181723356247, 0.15094339847564697, 0.19672130048274994, 0.22857142984867096, 0.20512819290161133, 0.060606054961681366, 0.043478257954120636, 0.1428571343421936, 0.060606054961681366, 0.1304347813129425, 0.2641509473323822, 0.17142856121063232, 0.17283950746059418, 0.10810810327529907, 0.09756097197532654 ]

[ "Systematically examines how well we can explain the hidden features of a deep network in terms of logical rules." ]

[ "Recent findings show that deep generative models can judge out-of-distribution samples as more likely than those drawn from the same distribution as the training data.", "In this work, we focus on variational autoencoders (VAEs) and address the problem of misaligned likelihood estimates on image data.", "We develop a novel likelihood function that is based not only on the parameters returned by the VAE but also on the features of the data learned in a self-supervised fashion.", "In this way, the model additionally captures the semantic information that is disregarded by the usual VAE likelihood function.", "We demonstrate the improvements in reliability of the estimates with experiments on the FashionMNIST and MNIST datasets.", "Deep Generative Models (DGMs) have gained in popularity due to their ability to model the density of the observed training data from which one can draw novel samples.", "However, as Nalisnick et al. (2018) pointed out in their recent paper, the inferences made by likelihood-based models, such as Variational Autoencoders (VAEs) (Kingma and Welling, 2015; Rezende et al., 2014) and flow-based models (Kingma and Dhariwal, 2018; van den Oord et al., 2016) , are not always reliable.", "They can judge out-of-distribution (OOD) samples to be more likely than in-distribution (ID) samples that are drawn from the same distribution as the training data.", "Concretely, a DGM trained on the FashionMNIST dataset will on average assign higher likelihoods to images from the MNIST dataset than to test images from the FashionMNIST dataset (see for example top left image in Figure 1(a) ).", "In this work we tackle the problem of misaligned likelihood estimates produced by VAEs on image data and propose a novel likelihood estimation during test time.", "Our method leverages findings reported in our earlier work Bütepage et al. (2019) , which are summarised in Section 2, and is based on the idea to evaluate a given test image not only locally, using individual parameters returned by a VAE as it is usually done, but also globally using learned feature representations of the data.", "The main contribution of this paper is the introduction of a feature-based likelihood trained in a self-supervised fashion.", "This likelihood evaluates the model also based on the semantics of a given image and not solely on the values of each pixel.", "We elaborate on this idea in Section 3 and demonstrate the improvements with an empirical evaluation presented in Section 4.", "We emphasise that the aim of our work is exclusively to improve the reliability of the likelihood estimation produced by VAEs.", "We focus on image data in particular as we have not observed the misalignment in our earlier experiments on various non-image datasets from UCI Machine Learning Repository (Dua and Graff, 2017) .", "We plan to investigate this further in the future work.", "Due to the lack of space we omit the experiments on non-image data as well as the specifics of VAEs for which we refer the reader to Kingma and Welling (2015) ; Rezende et al. (2014) .", "We have discussed how the problematic assumption that the image pixels are iid around the decoded parameters narrows the focus of the VAE likelihood function p V AE to a local area of the data density.", "Thus, the model likelihood function disregards the global data density, including the semantic information.", "Our proposed likelihood function mitigates this problem by leveraging self-supervised feature learning.", "In the future, we aim to evaluate our method on more complex datasets, such as CIFAR-10 and SVHN, and to design an end-to-end training procedure of VAEs using our proposed likelihood." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0 ]

[ 0, 0.27586206793785095, 0.1666666567325592, 0.07407406717538834, 0.1599999964237213, 0.0555555522441864, 0.039215683937072754, 0, 0.05128204822540283, 0.11428570747375488, 0.09677419066429138, 0.23076923191547394, 0.06896550953388214, 0.0714285671710968, 0.0714285671710968, 0.05128204822540283, 0.09999999403953552, 0, 0.04999999701976776, 0.09090908616781235, 0.3636363446712494, 0.10526315122842789 ]

[ "Improved likelihood estimates in variational autoencoders using self-supervised feature learning" ]

[ "Direct policy gradient methods for reinforcement learning and continuous control problems are a popular\napproach for a variety of reasons: \n1) they are easy to implement without explicit knowledge of the underlying model;\n2) they are an \"end-to-end\" approach, directly optimizing the performance metric of interest;\n3) they inherently allow for richly parameterized policies.\n", "A notable drawback is that even in the most basic continuous control problem (that of linear quadratic regulators), these methods must solve a non-convex optimization problem, where little is understood about their efficiency from both computational and statistical perspectives.", "In contrast, system identification and model based planning in optimal control theory have a much more solid theoretical footing, where much is known with regards to their computational and statistical properties. ", "This work bridges this gap showing that (model free) policy gradient methods globally converge to the optimal solution and are efficient (polynomially so in relevant problem dependent quantities) with regards to their sample and computational complexities.", "Recent years have seen major advances in the control of uncertain dynamical systems using reinforcement learning and data-driven approaches; examples range from allowing robots to perform more sophisticated controls tasks such as robotic hand manipulation (Tassa et al., 2012; BID1 Kumar et al., 2016; Levine et al., 2016; Tobin et al., 2017; Rajeswaran et al., 2017a) , to sequential decision making in game domains, e.g. AlphaGo (Silver et al., 2016) and Atari game playing (Mnih et al., 2015) .", "Deep reinforcement learning (DeepRL) are becoming increasingly popular for tackling such challenging sequential decision making problems.Many of these successes have relied on sampling based reinforcement learning algorithms such as policy gradient methods, including the DeepRL approaches; here, there is little theoretical understanding of their efficiency, either from a statistical or a computational perspective.", "In contrast, control theory (optimal and adaptive control) has a rich body of tools, with provable guarantees, for related sequential decision making problems, particularly those that involve continuous control.", "These latter techniques are often model-based -they estimate an explicit dynamical model first (e.g. system identification) and then design optimal controllers.", "This work builds bridges between these two lines of work, namely, between optimal control theory and sample based reinforcement learning methods, using ideas from mathematical optimization.", "This work has provided provable guarantees that model-based gradient methods and model-free (sample based) policy gradient methods convergence to the globally optimal solution, with finite polynomial computational and sample complexities.", "Taken together, the results herein place these popular and practical policy gradient approaches on a firm theoretical footing, making them comparable to other principled approaches (e.g. subspace ID methods and algebraic iterative approaches).Finite", "C(K 0 ) assumption, noisy case, and finite horizon case. These", "methods allow for extensions to the noisy case and the finite horizon case. This", "work also made the assumption that C(K 0 ) is finite, which may not be easy to achieve in some infinite horizon problems. The", "simplest way to address this is to model the infinite horizon problem with a finite horizon one; the techniques developed in Section D.1 shows this is possible. This", "is an important direction for future work.Open Problems.• Variance", "reduction: This work only proved efficiency from a polynomial sample size perspective. An interesting", "future direction would be in how to rigorously combine variance reduction methods and model-based methods to further decrease the sample size.• A sample based", "Gauss-Newton approach: This work showed how the Gauss-Newton algorithm improves over even the natural policy gradient method, in the exact case. A practically relevant", "question for the Gauss-Newton method would be how to both: a) construct a sample", "based estimator b) extend this scheme", "to deal with (non-linear) parametric policies.• Robust control: In model", "based approaches, optimal control theory provides efficient procedures to deal with (bounded) model mis-specification. An important question is how", "to provably understand robustness in a model free setting." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.2461538463830948, 0.17241378128528595, 0.11999999731779099, 0.37037035822868347, 0.0731707289814949, 0.14492753148078918, 0.1249999925494194, 0.0476190410554409, 0.13333332538604736, 0.3829787075519562, 0.18867923319339752, 0, 0.3125, 0.1818181723356247, 0.1818181723356247, 0.06451612710952759, 0.05882352590560913, 0.1428571343421936, 0.19512194395065308, 0.1764705777168274, 0, 0.06451612710952759, 0.1538461446762085, 0.06896550953388214 ]

[ "This paper shows that model-free policy gradient methods can converge to the global optimal solution for non-convex linearized control problems." ]

[ "Low-dimensional vector embeddings, computed using LSTMs or simpler techniques, are a popular approach for capturing the “meaning” of text and a form of unsupervised learning useful for downstream tasks.", "However, their power is not theoretically understood.", "The current paper derives formal understanding by looking at the subcase of linear embedding schemes.", "Using the theory of compressed sensing we show that representations combining the constituent word vectors are essentially information-preserving linear measurements of Bag-of-n-Grams (BonG) representations of text.", "This leads to a new theoretical result about LSTMs: low-dimensional embeddings derived from a low-memory LSTM are provably at least as powerful on classification tasks, up to small error, as a linear classifier over BonG vectors, a result that extensive empirical work has thus far been unable to show.", "Our experiments support these theoretical findings and establish strong, simple, and unsupervised baselines on standard benchmarks that in some cases are state of the art among word-level methods.", "We also show a surprising new property of embeddings such as GloVe and word2vec: they form a good sensing matrix for text that is more efficient than random matrices, the standard sparse recovery tool, which may explain why they lead to better representations in practice.", "Much attention has been paid to using LSTMs BID15 and similar models to compute text embeddings BID3 BID7 .", "Once trained, the LSTM can sweep once or twice through a given piece of text, process it using only limited memory, and output a vector with moderate dimensionality (a few hundred to a few thousand), which can be used to measure text similarity via cosine similarity or as a featurization for downstream tasks.The powers and limitations of this method have not been formally established.", "For example, can such neural embeddings compete with and replace traditional linear classifiers trained on trivial Bag-of-n-Grams (BonG) representations?", "Tweaked versions of BonG classifiers are known to be a surprisingly powerful baseline (Wang & Manning, 2012) and have fast implementations BID17 .", "They continue to give better performance on many downstream supervised tasks such as IMDB sentiment classification BID21 than purely unsupervised LSTM representations BID19 BID13 BID25 .", "Even a very successful character-level (and thus computation-intensive, taking a month of training) approach does not reach BonG performance on datasets larger than IMDB BID31 .", "Meanwhile there is evidence suggesting that simpler linear schemes give compact representations that provide most of the benefits of word-level LSTM embeddings (Wieting et al., 2016; BID1 .", "These linear schemes consist of simply adding up, with a few modifications, standard pretrained word embeddings such as GloVe or word2vec BID24 BID29 .The", "current paper ties these disparate threads together by giving an information-theoretic account of linear text embeddings. We", "describe linear schemes that preserve n-gram information as lowdimensional embeddings with provable guarantees for any text classification task. The", "previous linear schemes, which used unigram information, are subcases of our approach, but our best schemes can also capture n-gram information with low additional overhead. Furthermore", ", we show that the original unigram information can be (approximately) extracted from the low-dimensional embedding using sparse recovery/compressed sensing BID6 . Our approach", "also fits in the tradition of the older work on distributed representations of structured objects, especially the works of BID30 and BID18 . The following", "are the main results achieved by this new world-view:1. Using random", "vectors as word embeddings in our linear scheme (instead of pretrained vectors) already allows us to rigorously show that low-memory LSTMs are provably at least as good as every linear classifier operating on the full BonG vector. This is a novel", "theoretical result in deep learning, obtained relatively easily. By contrast, extensive", "empirical study of this issue has been inconclusive (apart from character-level models, and even then only on smaller datasets BID31 ). Note also that empirical", "work by its nature can only establish performance on some available datasets, not on all possible classification tasks. We prove this theorem in", "Section 4 by providing a nontrivial generalization of a result combining compressed sensing and learning BID5 ). In fact, before our work", "we do not know of any provable quantification of the power of any text embedding.2. We study theoretically and", "experimentally how our linear embedding scheme improves when it uses pretrained embeddings (GloVe etc.) instead of random vectors. Empirically we find that this", "improves the ability to preserve Bag-of-Words (BoW) information, which has the following restatement in the language of sparse recovery: word embeddings are better than random matrices for \"sensing\" BoW signals (see Section 5). We give some theoretical justification", "for this surprising finding using a new sparse recovery property characterizing when nonnegative signals can be reconstructed by 1 -minimization.3. Section 6 provides empirical results supporting", "the above theoretical work, reporting accuracy of our linear schemes on multiple standard classification tasks. Our embeddings are consistently competitive with", "recent results and perform much better than all previous linear methods. Among unsupervised word-level representations they", "achieve state of the art performance on both the binary and fine-grained SST sentiment classification tasks BID33 . Since our document representations are fast, compositional", ", and simple to implement given standard word embeddings, they provide strong baselines for future work.", "In this paper we explored the connection between compressed sensing, learning, and natural language representation.", "We first related LSTM and BonG methods via word embeddings, coming up with simple new document embeddings based on tensor product sketches.", "Then we studied their classification performance, proving a generalization of the compressed learning result of BID5 to convex Lipschitz losses and a bound on the loss of a low-dimensional LSTM classifier in terms of its (modified) BonG counterpart, an issue which neither experiments nor theory have been able to resolve.", "Finally, we showed how pretrained embeddings fit into this sparse recovery framework, demonstrating and explaining their ability to efficiently preserve natural language information.", "A COMPRESSED SENSING BACKGROUNDThe field of compressed sensing is concerned with recovering a high-dimensional k-sparse signal x ∈ R N from few linear measurements.", "In the noiseless case this is formulated as minimize w 0 subject to Aw = zwhere A ∈ R d×N is the design matrix and z = Ax is the measurement vector.", "Since 0 -minimization is NP-hard, a foundational approach is to use its convex surrogate, the 1 -norm, and characterize when the solution to (10) is equivalent to that of the following LP, known as basis pursuit (BP): DISPLAYFORM0 Related approaches such as Basis Pursuit Denoising (LASSO) and the Dantzig Selector generalize BP to handle signal or measurement noise BID11 ; however, the word embeddings case is noiseless so these methods reduce to BP.", "Note that throughout Section 5 and the Appendix we say that an 1 -minimization method recovers x from Ax if its optimal solution is unique and equivalent to the optimal solution of (10).An", "alternative way to approximately solve FORMULA1 is to use a greedy algorithm such as matching pursuit (MP) or orthogonal matching pursuit (OMP), which pick basis vectors one at a time by multiplying the measurement vector by A T and choosing the column with the largest inner product BID36 ." ]

[ 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.1666666567325592, 0, 0.21621620655059814, 0.40909090638160706, 0.25, 0.16326530277729034, 0.2461538463830948, 0.1538461446762085, 0.15584415197372437, 0.19512194395065308, 0.09090908616781235, 0.1702127605676651, 0.08695651590824127, 0.1666666567325592, 0.1304347813129425, 0.20512819290161133, 0.24390242993831635, 0.12765957415103912, 0.1818181723356247, 0.1428571343421936, 0.060606054961681366, 0.3333333432674408, 0, 0.1304347813129425, 0.22727271914482117, 0.1395348757505417, 0.25641024112701416, 0.1304347813129425, 0.13793103396892548, 0.04081632196903229, 0.23255813121795654, 0.052631575614213943, 0.17777776718139648, 0.052631575614213943, 0.10810810327529907, 0.09090908616781235, 0.2153846174478531, 0.04444443807005882, 0.17391303181648254, 0.12244897335767746, 0.14814814925193787, 0.15686273574829102, 0.158730149269104 ]

[ "We use the theory of compressed sensing to prove that LSTMs can do at least as well on linear text classification as Bag-of-n-Grams." ]

[ " Some conventional transforms such as Discrete Walsh-Hadamard Transform (DWHT) and Discrete Cosine Transform (DCT) have been widely used as feature extractors in image processing but rarely applied in neural networks.", "However, we found that these conventional transforms have the ability to capture the cross-channel correlations without any learnable parameters in DNNs.", "This paper firstly proposes to apply conventional transforms on pointwise convolution, showing that such transforms significantly reduce the computational complexity of neural networks without accuracy performance degradation.", "Especially for DWHT, it requires no floating point multiplications but only additions and subtractions, which can considerably reduce computation overheads.", "In addition, its fast algorithm further reduces complexity of floating point addition from O(n^2) to O(nlog n).", "These non-parametric and low computational properties construct extremely efficient networks in the number parameters and operations, enjoying accuracy gain.", "Our proposed DWHT-based model gained 1.49% accuracy increase with 79.4% reduced parameters and 48.4% reduced FLOPs compared with its baseline model (MoblieNet-V1) on the CIFAR 100 dataset.", "Large Convolutional Neural Networks (CNNs) (Krizhevsky et al., 2012; Simonyan & Zisserman, 2014; He et al., 2016; Szegedy et al., 2016b; a) and automatic Neural Architecture Search (NAS) based networks Liu et al., 2018; Real et al., 2018) have evolved to show remarkable accuracy on various tasks such as image classification (Deng et al., 2009; Krizhevsky & Hinton, 2009) , object detection (Lin et al., 2014) , benefited from huge amount of learnable parameters and computations.", "However, these large number of weights and high computational cost enabled only limited applications for mobile devices that require the constraint on memory space being low as well as for devices that require real-time computations (Canziani et al., 2016) .", "With regard to solving these problems, Howard et al. (2017) ; Sandler et al. (2018) ; Zhang et al. (2017b) ; Ma et al. (2018) proposed parameter and computation efficient blocks while maintaining almost same accuracy compared to other heavy CNN models.", "All of these blocks utilized depthwise separable convolution, which deconstructed the standard convolution with the (3 × 3 × C) size for each kernel into spatial information specific depthwise convolution (3 × 3 × 1) and channel information specific pointwise (1 × 1 × C) convolution.", "The depthwise separable convolution achieved comparable accuracy compared to standard spatial convolution with hugely reduced parameters and FLOPs.", "These reduced resource requirements made the depthwise separable convolution as well as pointwise convolution (PC) more widely used in modern CNN architectures.", "Nevertheless, we point out that the existing PC layer is still computationally expensive and occupies a lot of proportion in the number of weight parameters (Howard et al., 2017) .", "Although the demand toward PC layer has been and will be growing exponentially in modern neural network architectures, there has been a little research on improving the naive structure of itself.", "Therefore, this paper proposes a new PC layer formulated by non-parametric and extremely fast conventional transforms.", "Conventional transforms that we applied on CNN models are Discrete Walsh-Hadamard Transform (DWHT) and Discrete Cosine Transform (DCT), which have widely been used in image processing but rarely been applied in CNNs (Ghosh & Chellappa, 2016) .", "We empirically found that although both of these transforms do not require any learnable parameters at all, they show the sufficient ability to capture the cross-channel correlations.", "This non-parametric property enables our proposed CNN models to be significantly compressed in terms of the number of parameters, leading to get the advantages (i.e. efficient distributed training, less communication between server and clients) referred by Iandola et al. (2016) .", "We note that especially DWHT is considered to be a good replacement of the conventional PC layer, as it requires no floating point multiplications but only additions and subtractions by which the computation overheads of PC layers can significantly be reduced.", "Furthermore, DWHT can take a strong advantage of its fast version where the computation complexity of the floating point operations is reduced from O(n 2 ) to O(n log n).", "These non-parametric and low computational properties construct extremely efficient neural network from the perspective of parameter and computation as well as enjoying accuracy gain.", "Our contributions are summarized as follows:", "• We propose a new PC layer formulated with conventional transforms which do not require any learnable parameters as well as significantly reducing the number of floating point operations compared to the existing PC layer.", "• The great benefits of using the bases of existing transforms come from their fast versions, which drastically decrease computation complexity in neural networks without degrading accuracy.", "• We found that applying ReLU after conventional transforms discards important information extracted, leading to significant drop in accuracy.", "Based on this finding, we propose the optimal computation block for conventional transforms.", "• We also found that the conventional transforms can effectively be used especially for extracting high-level features in neural networks.", "Based on this, we propose a new transformbased neural network architecture.", "Specifically, using DWHT, our proposed method yields 1.49% accuracy gain as well as 79.4% and 49.4% reduced parameters and FLOPs, respectively, compared with its baseline model (MobileNet-V1) on CIFAR 100 dataset.", "We propose the new PC layers through conventional transforms.", "Our new PC layers allow the neural networks to be efficient in complexity of computation and learnable weight parameters.", "Especially for DWHT-based PC layer, its floating point multiplication-free property enabled extremely efficient in computation overhead.", "With the purpose of successfully fusing our PC layers into neural networks, we empirically found the optimal block unit structure and hierarchy level blocks in neural networks for conventional transforms, showing accuracy increase and great representability in cross-channel correlations.", "We further intrinsically revealed the hindrance of ReLU toward capturing the cross-channel representability and the activeness of depthwise convolution weights on the last blocks in our proposed neural network." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0 ]

[ 0.1904761791229248, 0.1666666567325592, 0.1904761791229248, 0, 0.060606054961681366, 0.11764705181121826, 0.0476190410554409, 0, 0, 0, 0.1249999925494194, 0.12121211737394333, 0.1666666567325592, 0.045454539358615875, 0.13636362552642822, 0.3125, 0.08510638028383255, 0.0952380895614624, 0.03703703358769417, 0.11320754140615463, 0.04651162400841713, 0.15789473056793213, 0, 0.21276594698429108, 0.1904761791229248, 0.22857142984867096, 0.13793103396892548, 0.277777761220932, 0.2222222238779068, 0.04255318641662598, 0.4000000059604645, 0.22857142984867096, 0.1249999925494194, 0.15686273574829102, 0.24390242993831635 ]

[ "We introduce new pointwise convolution layers equipped with extremely fast conventional transforms in deep neural network." ]

[ "We introduce the notion of \\emph{lattice representation learning}, in which the representation for some object of interest (e.g. a sentence or an image) is a lattice point in an Euclidean space.", "Our main contribution is a result for replacing an objective function which employs lattice quantization with an objective function in which quantization is absent, thus allowing optimization techniques based on gradient descent to apply; we call the resulting algorithms \\emph{dithered stochastic gradient descent} algorithms as they are designed explicitly to allow for an optimization procedure where only local information is employed.", "We also argue that a technique commonly used in Variational Auto-Encoders (Gaussian priors and Gaussian approximate posteriors) is tightly connected with the idea of lattice representations, as the quantization error in good high dimensional lattices can be modeled as a Gaussian distribution.", "We use a traditional encoder/decoder architecture to explore the idea of latticed valued representations, and provide experimental evidence of the potential of using lattice representations by modifying the \\texttt{OpenNMT-py} generic \\texttt{seq2seq} architecture so that it can implement not only Gaussian dithering of representations, but also the well known straight-through estimator and its application to vector quantization. \n", "The present work is inspired by a belief that information theory, and in particular lossy compression theory can be very effective in serving as a theoretical foundation for problems in representation learning, including the design and analysis of highly performant practical algorithms.", "We have introduced lattices as a possible way to create discrete representations, and proved a fundamental result which allows us to train computational networks that use lattice quantized dithering using an equivalent (in an expected sense) computational network which replaces quantization with dithering, thus allowing gradient descent to apply.", "This result also allows us to use only local information during the optimization, thus additionally enabling stochastic gradient descent.", "We also established a fundamental connection between the use of good high dimensional lattices and the idea of Gaussian dithering, which is common in generative modeling settings such as Variational Autoencoders.", "Finally, we provided initial experimental evidence of the potential of using lattices in an VAE setting, where we contrasted the performance of a rectangular lattice based VAE and two types of Gaussian VAEs.", "The bottom line is that if one is interested in getting close to the performance of a Gaussian VAE with discrete representations with a good theoretical basis, we suggest the reader to consider lattices and to train them using dithered stochastic gradient descent." ]

[ 0, 0, 0, 0, 0, 1, 0, 0, 0, 0 ]

[ 0.09090908616781235, 0.12121211737394333, 0.1818181723356247, 0.1846153736114502, 0.1071428507566452, 0.3606557250022888, 0.1621621549129486, 0.25531914830207825, 0.13333332538604736, 0.2545454502105713 ]

[ "We propose to use lattices to represent objects and prove a fundamental result on how to train networks that use them." ]

[ "There were many attempts to explain the trade-off between accuracy and adversarial robustness.", "However, there was no clear understanding of the behaviors of a robust classifier which has human-like robustness.\n", "We argue (1) why we need to consider adversarial robustness against varying magnitudes of perturbations not only focusing on a fixed perturbation threshold, (2) why we need to use different method to generate adversarially perturbed samples that can be used to train a robust classifier and measure the robustness of classifiers and (3) why we need to prioritize adversarial accuracies with different magnitudes.\n", "We introduce Lexicographical Genuine Robustness (LGR) of classifiers that combines the above requirements. ", "We also suggest a candidate oracle classifier called \"Optimal Lexicographically Genuinely Robust Classifier (OLGRC)\" that prioritizes accuracy on meaningful adversarially perturbed examples generated by smaller magnitude perturbations. ", "The training algorithm for estimating OLGRC requires lexicographical optimization unlike existing adversarial training methods.", "To apply lexicographical optimization to neural network, we utilize Gradient Episodic Memory (GEM) which was originally developed for continual learning by preventing catastrophic forgetting.", "Even though deep learning models have shown promising performances in image classification tasks [6] , most deep learning classifiers mis-classify imperceptibly perturbed images, i.e. adversarial examples [7] .", "This vulnerability can occur even when the adversarial attacks were applied before they print the images, and the printed images were read through a camera [8] .", "That result shows real-world threats of classifiers can exist.", "In addition, adversarial examples for a classifier can be transferable to other models [3] .", "This transferability of adversarial examples [9] enables attackers to exploit a target model with limited access to the target classifier.", "This kinds of attacks is called black-box attacks.", "In this work, we explained why existing adversarial training methods cannot train a classifier that has human-like robustness.", "We identified three properties of human-like classification: (1) human-like classification should be robust against varying magnitudes of adversarially perturbed samples and not just on a fixed maximum norm perturbations, (2) when we consider robustness on increasing magnitudes of adversarial perturbations, a human-like classifier should avoid considering already considered points multiple times, and (3) human-like classification need to prioritize the robustness against adversarially perturbed samples with smaller perturbation norm.", "The suggested properties explain why previous methods for adversarial training and evaluation can be incomplete.", "For example, the second property explains why commonly used evaluation of adversarial robustness may not fully reveal our intuitive understanding of human-like robustness as standard adversarial accuracies don't avoid pseudo adversarial examples.", "We defined a candidate oracle classifier called Optimal Lexicographically Genuinely Robust Classifier (OL-GRC).", "OLGRC is (almost everywhere) uniquely determined when dataset and norm were given.", "In order to train a OLGRC, we suggested a method to generate adversarially perturbed samples using a discriminator.", "We proposed to use Gradient Episodic Memory (GEM) [4] for lexicographical optimization [2] and an approach to applying GEM when simultaneously reducing multiple losses with lexicographical preferences.", "From the first experiment on the toy example from section 2, we showed that lexicographical optimization enables stable training even when other adversarial training methods failed to do so.", "The second experiment on the same toy example showed that we can use discriminator to roughly generate adversarially perturbed samples by avoiding already explored regions.", "Because of that, we could train a classifier that is similar to the theoretical OLGRC.", "From the experiment on the MNIST data, we showed that our methods (OLSRC and OLGRC) achieved better performances on natural accuracy and adversarial accuracy than using standard adversarial training method [3] ." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0 ]

[ 0.27586206793785095, 0.24242423474788666, 0.2857142686843872, 0.13333332538604736, 0.13636362552642822, 0.06896550953388214, 0.04999999701976776, 0.0476190410554409, 0.1538461446762085, 0.07999999821186066, 0.2666666507720947, 0.29411762952804565, 0.17391303181648254, 0.29411762952804565, 0.24242423474788666, 0.12903225421905518, 0.13636362552642822, 0.20689654350280762, 0.1428571343421936, 0.19354838132858276, 0.1463414579629898, 0.09302324801683426, 0.04878048226237297, 0.3870967626571655, 0.0952380895614624 ]

[ "We try to design and train a classifier whose adversarial robustness is more resemblance to robustness of human." ]

[ "Generating complex discrete distributions remains as one of the challenging problems in machine learning.", "Existing techniques for generating complex distributions with high degrees of freedom depend on standard generative models like Generative Adversarial Networks (GAN), Wasserstein GAN, and associated variations.", "Such models are based on an optimization involving the distance between two continuous distributions.", "We introduce a Discrete Wasserstein GAN (DWGAN) model which is based on a dual formulation of the Wasserstein distance between two discrete distributions.", "We derive a novel training algorithm and corresponding network architecture based on the formulation.", "Experimental results are provided for both synthetic discrete data, and real discretized data from MNIST handwritten digits.", "Generative Adversarial Networks (GAN) BID3 have gained significant attention in the field of machine learning.", "The goal of GAN models is to learn how to generate data based on a collection of training samples.", "The GAN provides a unique training procedure by treating the learning optimization as a two player game between a generator network and discriminator network.", "Since the learning process involves optimization over two different networks simultaneously, the GAN is hard to train, often times unstable BID11 .", "Newly developed models such as the Wasserstein GAN aim to improve the training process by leveraging the Wasserstein distance in optimization, as opposed to the Kullback-Leibler or Jensen-Shannon divergences utilized by the original GAN.A source of interest in generative models arises from natural language processing.", "In natural language applications, a generative model is necessary to learn complex distributions of text documents.", "Although both the GAN and Wasserstein GAN approximate a distance between two continuous distributions, and use a continuous sample distance, prior research efforts BID4 BID12 BID10 have applied the models to discrete probability distributions advocating for a few modifications.", "However, using a continuous sample distance for the discrete case may lead to discrepancies.", "More precisely, as will be demonstrated via explicit examples, a small continuous distance does not necessarily imply a small discrete distance.", "This observation has potentially serious ramifications for generating accurate natural language text and sentences using GAN models.To address the above issues, we propose a Discrete Wasserstein GAN (DWGAN) which is directly based on a dual formulation of the Wasserstein distance between two discrete distributions.", "A principal challenge is to enforce the dual constraints in the corresponding optimization.", "We derive a novel training algorithm and corresponding network architecture as one possible solution.", "We proposed the Discrete Wasserstein GAN (DWGAN) which approximates the Wasserstein distance between two discrete distributions.", "We derived a novel training algorithm and corresponding network architecture for a dual formulation to the problem, and presented promising experimental results.", "Our future work focuses on exploring techniques to improve the stability of the training process, and applying our model to other datasets such as for natural language processing." ]

[ 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.22857142984867096, 0.1702127605676651, 0.4000000059604645, 0.9523809552192688, 0.34285715222358704, 0.052631575614213943, 0.1111111044883728, 0.31578946113586426, 0.2380952388048172, 0.19512194395065308, 0.178571417927742, 0.2702702581882477, 0.3333333432674408, 0.22857142984867096, 0.1538461446762085, 0.6129032373428345, 0.1818181723356247, 0.11428570747375488, 0.6857143044471741, 0.24390242993831635, 0.1702127605676651 ]

[ "We propose a Discrete Wasserstein GAN (DWGAN) model which is based on a dual formulation of the Wasserstein distance between two discrete distributions." ]

[ "We introduce the open-ended, modular, self-improving Omega AI unification architecture which is a refinement of Solomonoff's Alpha architecture, as considered from first principles.", "The architecture embodies several crucial principles of general intelligence including diversity of representations, diversity of data types, integrated memory, modularity, and higher-order cognition.", "We retain the basic design of a fundamental algorithmic substrate called an ``AI kernel'' for problem solving and basic cognitive functions like memory, and a larger, modular architecture that re-uses the kernel in many ways.", "Omega includes eight representation languages and six classes of neural networks, which are briefly introduced.", "The architecture is intended to initially address data science automation, hence it includes many problem solving methods for statistical tasks.", "We review the broad software architecture, higher-order cognition, self-improvement, modular neural architectures, intelligent agents, the process and memory hierarchy, hardware abstraction, peer-to-peer computing, and data abstraction facility.", "In today's AI research, most researchers focus on specific application problems and they develop the capabilities of their AI solutions only to the extent that these specific applications require them.", "While challenging AI problems such as natural language understanding require a broader view, most researchers do not begin with an all-encompassing architecture and then adapt to a specific application.", "It is usually more efficient to pursue a bottom-up development methodology for the experimental results, and as a result, progress in ambitious architectures for generality may have stalled.To achieve generality, a rigorous architectural approach has several benefits such as easing development, allowing future extensions while remaining backwards compatible, and exposing problems before they happen since we can conceptualize complex use-cases.", "In other words, it is at least better software engineering, however, there are also scientific benefits such as understanding the functions and capabilities required by a general-purpose AI system much better, and address these problems fully.", "Since the most general problem is attacked, the architecture can follow a rigorous design process which will eliminate redundancies, leading us to a more mathematically elegant design.", "And finally, since use-cases will lead the design, the result will be empirically firmer than a special-purpose application.A design from first principles is rarely undertaken, and it is arduous, but it can produce highly effective systems.", "We build upon the most powerful architectures for general AI, and then identify the requirements, from which we introduce refinements to the existing architectures, introducing new architectural ideas and incorporating new AI technologies in the process.", "The resulting deep technological integration architecture is a compact, scalable, portable, AI platform for general-purpose AI with many possible applications in wide domains.", "We gave the overview of an ambitious architecture based on Solomonoff's Alpha Architecture, and Schmidhuber's Gödel Machine architecture.", "The system is like Alpha, because it re-uses the basic design of PSMs.", "It is also similar to Gödel Machine architecture, because it can deploy a kind of probabilistic logical inference for reasoning and it can also observe some of its internal states and improve itself.", "The system also has basic provisions for intelligent agents, but it is not limited to them.", "We saw that the first important issue with implementing Alpha was to decide a basic set of primitives that will grant it sufficient intelligence to deal with human-scale problems.", "It remains to be demonstrated empirically that is the case, however, two of the eight reference machines have been implemented and seen to operate effectively.A criticism may be raised that we have not explained much about how the AI Kernel works.", "We only assume that it presents a generalized universal induction approximation that can optimize functions, rich enough to let us define basic machine learning tasks.", "It surely cannot be Levin search, but it could be any effective multi-strategy optimization method such as evolutionary architecture search BID7 .", "We are using an extension of the approach in Fourier Network Search BID6 which is also likely general enough.", "The memory update is also not detailed but it is assumed that it is possible to extend an older memory design called heuristic algorithmic memory so that it works for any reference machine.", "We also did not explain in detail how many components work due to lack of space, which is an issue to be tackled in a longer future version of the present paper.In the future, we would like to support the architectural design with experiments, showing if the system is imaginative enough to come up with neural architectures or hybrid solutions that did not appear to humans.", "The algorithms used are expensive, therefore they might not work very well with the extremely large models required by the best vision processing systems; but to accommodate such models, it might be required that the system evolves only parts of the system and not the entire architecture.", "The system is intended to be tested on basic psychometric tests first, and a variety of data science problems to see if we can match the competence of the solution a human data scientist would achieve." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.2916666567325592, 0.13333332538604736, 0.178571417927742, 0.14999999105930328, 0.2666666507720947, 0.07999999821186066, 0.11538460850715637, 0.11320754140615463, 0.12345678359270096, 0.13333332538604736, 0.2448979616165161, 0.10344827175140381, 0.178571417927742, 0.1702127605676651, 0.1904761791229248, 0.21052631735801697, 0.18867923319339752, 0.19512194395065308, 0.15686273574829102, 0.13114753365516663, 0.08163265138864517, 0.04444443807005882, 0.1818181723356247, 0.11764705181121826, 0.17721518874168396, 0.15625, 0.2857142686843872 ]

[ "It's a new AGI architecture for trans-sapient performance.This is a high-level overview of the Omega AGI architecture which is the basis of a data science automation system. Submitted to a workshop. " ]

[ "Conversational machine comprehension requires a deep understanding of the conversation history.", "To enable traditional, single-turn models to encode the history comprehensively, we introduce Flow, a mechanism that can incorporate intermediate representations generated during the process of answering previous questions, through an alternating parallel processing structure.", "Compared to shallow approaches that concatenate previous questions/answers as input, Flow integrates the latent semantics of the conversation history more deeply.", "Our model, FlowQA, shows superior performance on two recently proposed conversational challenges (+7.2% F1 on CoQA and +4.0% on QuAC).", "The effectiveness of Flow also shows in other tasks.", "By reducing sequential instruction understanding to conversational machine comprehension, FlowQA outperforms the best models on all three domains in SCONE, with +1.8% to +4.4% improvement in accuracy.", "Q3", "We presented a novel FLOW component for conversational machine comprehension.", "By applying FLOW to a state-of-the-art machine comprehension model, our model encodes the conversation history more comprehensively, and thus yields better performance.", "When evaluated on two recently proposed conversational challenge datasets and three domains of a sequential instruction understanding task (through reduction), FLOWQA outperforms existing models.While our approach provides a substantial performance gain, there is still room for improvement.", "In the future, we would like to investigate more efficient and fine-grained ways to model the conversation flow, as well as methods that enable machines to engage more active and natural conversational behaviors, such as asking clarification questions." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0 ]

[ 0.1764705777168274, 0.178571417927742, 0.1395348757505417, 0.23255813121795654, 0.0624999962747097, 0.23999999463558197, 0.1818181723356247, 0.13333332538604736, 0.3333333432674408, 0.14814814925193787 ]

[ "We propose the Flow mechanism and an end-to-end architecture, FlowQA, that achieves SotA on two conversational QA datasets and a sequential instruction understanding task." ]

[ "We consider reinforcement learning and bandit structured prediction problems with very sparse loss feedback: only at the end of an episode.", "We introduce a novel algorithm, RESIDUAL LOSS PREDICTION (RESLOPE), that solves such problems by automatically learning an internal representation of a denser reward function.", "RESLOPE operates as a reduction to contextual bandits, using its learned loss representation to solve the credit assignment problem, and a contextual bandit oracle to trade-off exploration and exploitation.", "RESLOPE enjoys a no-regret reduction-style theoretical guarantee and outperforms state of the art reinforcement learning algorithms in both MDP environments and bandit structured prediction settings.", "Current state of the art learning-based systems require enormous, costly datasets on which to train supervised models.", "To progress beyond this requirement, we need learning systems that can interact with their environments, collect feedback (a loss or reward), and improve continually over time.", "In most real-world settings, such feedback is sparse and delayed: most decisions made by the system will not immediately lead to feedback.", "Any sort of interactive system like this will face at least two challenges: the credit assignment problem (which decision(s) did the system make that led to the good/bad feedback?", "); and the exploration/exploitation problem (in order to learn, the system must try new things, but these could be bad).We", "consider the question of how to learn in an extremely sparse feedback setting: the environment operates episodically, and the only feedback comes at the end of the episode, with no incremental feedback to guide learning. This", "setting naturally arises in many classic reinforcement learning problems ( §4): a barista robot will only get feedback from a customer after their cappuccino is finished 1 . It also", "arises in the context of bandit structured prediction BID41 BID9 ( §2.2), where a structured prediction system must produce a single output (e.g., translation) and observes only a scalar loss.We introduce a novel reinforcement learning algorithm, RESIDUAL LOSS PREDICTION (RESLOPE) ( § 3), which aims to learn effective representations of the loss signal. By effective", "we mean effective in terms of credit assignment. Intuitively,", "RESLOPE attempts to learn a decomposition of the episodic loss into a sum of per-time-step losses. This process", "is akin to how a person solving a task might realize before the task is complete when and where they are likely to have made suboptimal choices. In RESLOPE,", "the per-step loss estimates are conditioned on all the information available up to the current point in time, allowing it to learn a highly non-linear representation for the episodic loss (assuming the policy class is sufficiently complex; in practice, we use recurrent neural network policies). When the system", "receives the final episodic loss, it uses the difference between the observed loss and the cumulative predicted loss to update its parameters.Algorithmically, RESLOPE operates as a reduction ( §3.3) to contextual bandits (Langford & Zhang, 2008) , allowing the bandit algorithm to handle exploration/exploitation and focusing only on the credit assignment problem. RESIDUAL LOSS PREDICTION", "is theoretically motivated by the need for variance reduction techniques when estimating counterfactual costs (Dudík et al., 2014) and enjoys a no-regret bound ( §3.3) when the underlying bandit algorithm is no-regret. Experimentally, we show", "the efficacy of RESLOPE on four benchmark reinforcement problems and three bandit structured prediction problems ( § 5.1), comparing to several reinforcement learning algorithms: Reinforce, Proximal Policy Optimization and Advantage Actor-Critic.", "RESIDUAL LOSS PREDICTION builds most directly on the bandit learning to search frameworks LOLS BID9 and BLS BID40 .", "The \"bandit\" version of LOLS was analyzed theoretically but not empirically in the original paper; BID40 found that it failed to learn empirically.They addressed this by requiring additional feedback from the user, which worked well empirically but did not enjoy any theoretical guarantees.", "RESLOPE achieves the best of both worlds: a strong regret guarantee, good empirical performance, and no need for additional feedback.", "The key ingredient for making this work is using the residual loss structure together with strong base contextual bandit learning algorithms.A number of recent algorithms have updated \"classic\" learning to search papers with deep learning underpinnings BID48 BID21 .", "These aim to incorporate sequencelevel global loss function to mitigate the mismatch between training and test time discrepancies, but only apply in the fully supervised setting.", "Mixing of supervised learning and reinforcement signals has become more popular in structured prediction recently, generally to do a better job of tuning for a task-specific loss using either Reinforce BID35 or Actor-Critic BID2 .", "The bandit variant of the structured prediction problem was studied by BID41 , who proposed a reinforce method for optimizing different structured prediction models under bandit feedback in a log-linear structured prediction model.A standard technique for dealing with sparse and episodic reward signals is reward shaping BID31 : supplying additional rewards to a learning agent to guide its learning process, beyond those supplied by the underlying environment.", "Typical reward shaping is hand-engineered; RESLOPE essentially learns a good task-specific reward shaping automatically.", "The most successful baseline approach we found is Proximal Policy Optimization (PPO, BID39 ), a variant of Trust Region Policy Optimization (TRPO, BID38 ) that is more practical.Experimentally we have seen RESLOPE to typically learn more quickly than PPO.", "Theoretically both have useful guarantees of a rather incomparable nature.Since RESLOPE operates as a reduction to a contextual bandit oracle, this allows it to continually improve as better contextual bandit algorithms become available, for instance work of Syrgkanis et al. (2016b) and BID0 .", "Although RESLOPE is quite effective, there are a number of shortcomings that need to be addressed in future work.", "For example, the bootstrap sampling algorithm is prohibitive in terms of both memory and time efficiency.", "One approach for tackling this would be using the amortized bootstrap approach by BID27 , which uses amortized inference in conjunction with implicit models to approximate the bootstrap distribution over model parameters.", "There is also a question of whether the reduction to contextual bandits creates \"reasonable\" contextual bandit problems in conjunction with RNNs.", "While some contextual bandit algorithms assume strong convexity or linearity, the ones we employ operate on arbitrary policy classes, provided a good cost-sensitive learner exists.", "The degree to which this is true will vary by neural network architecture, and what can be guaranteed (e.g., no regret full-information online neural learning).", "A more significant problem in the multi-deviation setting is that as RESLOPE learns, the residual costs will change, leading to a shifting distribution of costs; in principle this could be addressed using CB algorithms that work in adversarial settings BID43 BID16 , but largely remains an open challenge.", "RESLOPE is currently designed for discrete action spaces.", "Extension to continuous action spaces BID22 BID23 remains an open problem." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.6000000238418579, 0.2380952388048172, 0.1860465109348297, 0.3255814015865326, 0, 0.2222222238779068, 0.1538461446762085, 0, 0.10256409645080566, 0.21276594698429108, 0.21276594698429108, 0.29411762952804565, 0, 0.11428570747375488, 0.13636362552642822, 0.10169491171836853, 0.1492537260055542, 0.19230768084526062, 0.2916666567325592, 0.1621621549129486, 0.03448275476694107, 0.20512819290161133, 0.18518517911434174, 0.09302324801683426, 0.31372547149658203, 0.2702702581882477, 0.06451612710952759, 0.03703703358769417, 0.1428571343421936, 0.052631575614213943, 0.11428570747375488, 0.08510638028383255, 0.20512819290161133, 0.09090908616781235, 0.04444443807005882, 0.0317460261285305, 0.07407406717538834, 0 ]

[ "We present a novel algorithm for solving reinforcement learning and bandit structured prediction problems with very sparse loss feedback." ]

[ "Adversarial neural networks solve many important problems in data science, but are notoriously difficult to train.", "These difficulties come from the fact that optimal weights for adversarial nets correspond to saddle points, and not minimizers, of the loss function.", "The alternating stochastic gradient methods typically used for such problems do not reliably converge to saddle points, and when convergence does happen it is often highly sensitive to learning rates.", "We propose a simple modification of stochastic gradient descent that stabilizes adversarial networks.", "We show, both in theory and practice, that the proposed method reliably converges to saddle points.", "This makes adversarial networks less likely to \"collapse,\" and enables faster training with larger learning rates.", "Adversarial networks play an important role in a variety of applications, including image generation (Zhang et al., 2017; Wang & Gupta, 2016) , style transfer BID2 Taigman et al., 2017; Wang & Gupta, 2016; BID17 , domain adaptation (Taigman et al., 2017; Tzeng et al., 2017; BID11 , imitation learning BID15 , privacy BID9 BID0 , fair representation (Mathieu et al., 2016; BID9 , etc. One particularly motivating application of adversarial nets is their ability to form generative models, as opposed to the classical discriminative models BID13 Radford et al., 2016; BID7 Mirza & Osindero, 2014) .While", "adversarial networks have the power to attack a wide range of previously unsolved problems, they suffer from a major flaw: they are difficult to train. This", "is because adversarial nets try to accomplish two objectives simultaneously; weights are adjusted to maximize performance on one task while minimizing performance on another. Mathematically", ", this corresponds to finding a saddle point of a loss function -a point that is minimal with respect to one set of weights, and maximal with respect to another.Conventional neural networks are trained by marching down a loss function until a minimizer is reached ( FIG0 ). In contrast,", "adversarial training methods search for saddle points rather than a minimizer, which introduces the possibility that the training path \"slides off\" the objective functions and the loss goes to −∞ FIG0 ), resulting in \"collapse\" of the adversarial network. As a result,", "many authors suggest using early stopping, gradients/weight clipping , or specialized objective functions BID13 Zhao et al., 2017; to maintain stability.In this paper, we present a simple \"prediction\" step that is easily added to many training algorithms for adversarial nets. We present theoretical", "analysis showing that the proposed prediction method is asymptotically stable for a class of saddle point problems. Finally, we use a wide", "range of experiments to show that prediction enables faster training of adversarial networks using large learning rates without the instability problems that plague conventional training schemes. If minimization (or, conversely", ", maximization) is more powerful, the solution path \"slides off\" the loss surface and the algorithm becomes unstable, resulting in a sudden \"collapse\" of the network.", "We present a simple modification to the alternating SGD method, called a prediction step, that improves the stability of adversarial networks.", "We present theoretical results showing that the prediction step is asymptotically stable for solving saddle point problems.", "We show, using a variety of test problems, that prediction steps prevent network collapse and enable training with a wider range of learning rates than plain SGD methods." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0 ]

[ 0.11428570747375488, 0.24390242993831635, 0.0833333283662796, 0.5, 0.22857142984867096, 0.17142856121063232, 0.13483145833015442, 0.2857142686843872, 0.09756097197532654, 0.17241378128528595, 0.2222222238779068, 0.2666666507720947, 0.25, 0.30434781312942505, 0.1428571343421936, 1, 0.277777761220932, 0.2666666507720947 ]

[ "We present a simple modification to the alternating SGD method, called a prediction step, that improves the stability of adversarial networks." ]

[ "An important type of question that arises in Explainable Planning is a contrastive question, of the form \"Why action A instead of action B?\".", "These kinds of questions can be answered with a contrastive explanation that compares properties of the original plan containing A against the contrastive plan containing B. An effective explanation of this type serves to highlight the differences between the decisions that have been made by the planner and what the user would expect, as well as to provide further insight into the model and the planning process.", "Producing this kind of explanation requires the generation of the contrastive plan.", "This paper introduces domain-independent compilations of user questions into constraints.", "These constraints are added to the planning model, so that a solution to the new model represents the contrastive plan.", "We introduce a formal description of the compilation from user question to constraints in a temporal and numeric PDDL2.1 planning setting.", "Explainable AI (XAI) is an emerging and important research area within AI.", "Recent work has shown that AI Planning is an important tool in XAI, as its decision-making mechanisms are model-based and so in principle more transparent.", "This recent work includes many approaches towards providing explanations in AI planning.", "BID3 gives an in-depth overview of this work and different terms used within the XAI landscape.", "In particular, BID16 shows that if an AI system behaves \"explicably\" there is less of a need for explanations.", "However, this is not always possible and explanation is sometimes required.", "BID2 tackles explanation as a model reconciliation problem, arguing that the explanation must be a difference between the human model and AI model.", "BID14 show that by representing plans as first order logic formulae generating explanations is feasible in real time.", "In contrast, in this paper we focus on contrastive \"why\" questions.", "BID4 highlight some important questions in XAIP and discuss possible answers, and also describe how these \"why\" questions are especially important.", "BID15 outlines the approach to planning as an iterative process for bet- ter modelling preferences and providing explanations.", "We propose to follow this same approach.The aim of explanations is to improve the user's levels of understanding and trust in the system they are using.", "These explanations can be local (regarding a specific plan) or global (concerning how the planning system works in general).", "In this paper we focus on local explanations of temporal and numeric planning problems, introducing an approach for explaining why a planner has made a certain decision.", "Through active exploration of these specific cases, the user may also gain global insight into the way in which the planner makes decisions.", "(See BID9 BID10 Ribeiro, Singh, and Guestrin 2016) ).To", "achieve an understanding of a decision, it is important that explanations adapt to the specific context and mental model of the user. One", "step towards this is to support the user iteratively asking different questions suitable for their context. BID6", "identify ten question types that a user might have about an intelligent system, also described by BID13 . BID8", "show in a grounded study that of these, the questions why and why not provided the most benefit in terms of objective understanding and feelings of trust. In the", "context of planning why not questions are contrastive questions, because the user is asking why some action was selected rather than some other action that was not.Instead, Miller argues that all such questions can be asked as contrastive questions of the form \"Why action A rather than action B?\" BID11", ". Contrastive", "questions capture the context of the question; they more precisely identify the gaps in the user's understanding of a plan that needs to be explained BID7 . A contrastive", "question about a plan can be answered by a contrastive explanation. Contrastive explanations", "will compare the original plan against a contrastive plan that accounts for the user expectation. Providing contrastive explanations", "is not only effective in improving understanding, but is simpler than providing a full causal analysis BID12 .Following the approach of Smith (2012", ") we propose an approach to contrastive explanations through a dialogue with the user. The proposed approach consists of an", "iterative four-stage process illustrated in FIG0 . First the user asks a contrastive question", "in natural language. Second, a constraint is derived from the user", "question, in the following we refer to this constraint as the formal question. Third a hypothetical model (HModel) is generated", "which encapsulates this constraint. A solution to this model is the hypothetical plan", "(HPlan) that can be compared to the original plan to show the consequence of the user suggestion. The user can compare plans and iterate the process", "by asking further questions, and refining the HModel. This allows the user to combine different compilations", "to create a more constrained HModel, producing more meaningful explanations, until the explanation is satisfactory. Each stage of this process represents a vital research", "challenge. This paper describes and formalises the third stage of", "this process: compiling the formal question into a hypothetical model for temporal and numeric planning.We are interested in temporal and numeric planning problems, for which optimal solutions are difficult to find. Therefore, while the process described above serves for", "explanation, the insight of the user can also result in guiding the planning process to a more efficient solution. As noted by BID15 , the explanations could also give the", "user the opportunity to improve the plan with respect to their own preferences. The user could have hidden preferences which have not been", "captured in the model. The user could ask questions which enforce constraints that", "favour these preferences. The new plan could be sub-optimal, but more preferable to the", "user.The contribution of this paper is a formalisation of domain-independent and planner-agnostic compilations from formal contrastive questions to PDDL2.1 (Fox and Long 2003) , necessary for providing contrastive explanations. The compilations shown are not exhaustive. However, they do cover", "an interesting set of questions which users", "would commonly have about both classical and temporal plans. The paper is organised as follows. The next section describes the", "planning definitions we will use throughout", "the paper. In Section 3 we describe the running example that we use to demonstrate our", "compilations throughout the paper. In Section 4 we list the set of formal questions that we are interested in,", "and formalise the compilations of each of these into constraints. Finally, we conclude the paper in Section 5 whilst touching on some interesting", "future work." ]

[ 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.11764705181121826, 0.158730149269104, 0.17391303181648254, 0.8695651888847351, 0.13333332538604736, 0.1764705777168274, 0, 0, 0.1599999964237213, 0.06896550953388214, 0.1875, 0, 0, 0.06451612710952759, 0.25, 0.06451612710952759, 0.12903225421905518, 0.10810810327529907, 0.0624999962747097, 0.20512820780277252, 0.1764705777168274, 0, 0.1764705777168274, 0.19999998807907104, 0.06451612710952759, 0.11428570747375488, 0.16326530277729034, 0.1621621549129486, 0.1599999964237213, 0.2857142686843872, 0.05714285373687744, 0.25806450843811035, 0.1538461446762085, 0.0833333283662796, 0, 0, 0.12121211737394333, 0.2142857164144516, 0.05714285373687744, 0.260869562625885, 0.08888888359069824, 0.15789473056793213, 0.06451612710952759, 0.23076923191547394, 0, 0.36734694242477417, 0.19999998807907104, 0.0624999962747097, 0, 0.07407406717538834, 0.2666666507720947, 0.29411762952804565 ]

[ "This paper introduces domain-independent compilations of user questions into constraints for contrastive explanations." ]

[ "Methods that learn representations of nodes in a graph play a critical role in network analysis since they enable many downstream learning tasks.", "We propose Graph2Gauss - an approach that can efficiently learn versatile node embeddings on large scale (attributed) graphs that show strong performance on tasks such as link prediction and node classification.", "Unlike most approaches that represent nodes as point vectors in a low-dimensional continuous space, we embed each node as a Gaussian distribution, allowing us to capture uncertainty about the representation.", "Furthermore, we propose an unsupervised method that handles inductive learning scenarios and is applicable to different types of graphs: plain/attributed, directed/undirected.", "By leveraging both the network structure and the associated node attributes, we are able to generalize to unseen nodes without additional training.", "To learn the embeddings we adopt a personalized ranking formulation w.r.t. the node distances that exploits the natural ordering of the nodes imposed by the network structure.", "Experiments on real world networks demonstrate the high performance of our approach, outperforming state-of-the-art network embedding methods on several different tasks.", "Additionally, we demonstrate the benefits of modeling uncertainty - by analyzing it we can estimate neighborhood diversity and detect the intrinsic latent dimensionality of a graph.", "Graphs are a natural representation for a wide variety of real-life data, from social and rating networks (Facebook, Amazon), to gene interactions and citation networks (BioGRID, arXiv) .", "Node embeddings are a powerful and increasingly popular approach to analyze such data BID0 .", "By operating in the embedding space, one can employ proved learning techniques and bypass the difficulty of incorporating the complex node interactions.", "Tasks such as link prediction, node classification, community detection, and visualization all greatly benefit from these latent node representations.", "Furthermore, for attributed graphs by leveraging both sources of information (network structure and attributes) one is able to learn more useful representations compared to approaches that only consider the graph BID33 BID24 BID5 .All", "existing (attributed) graph embedding approaches represent each node by a single point in a low-dimensional continuous vector space. Representing", "the nodes simply as points, however, has a crucial limitation: we do not have information about the uncertainty of that representation. Yet uncertainty", "is inherent when describing a node in a complex graph by a single point only. Imagine a node", "for which the different sources of information are conflicting with each other, e.g. pointing to different communities or even revealing contradicting underlying patterns. Such discrepancy", "should be reflected in the uncertainty of its embedding. As a solution to", "this problem, we introduce a novel embedding approach that represents nodes as Gaussian distributions: each node becomes a full distribution rather than a single point. Thereby, we capture", "uncertainty about its representation.To effectively capture the non-i.i.d. nature of the data arising from the complex interactions between the nodes, we further propose a novel unsupervised personalized ranking formulation to learn the embeddings. Intuitively, from the", "point of view of a single node, we want nodes in its immediate neighborhood to be closest in the embedding space, while nodes multiple hops away should become increasingly more distant. This ordering between", "the nodes imposed by the network structure w.r.t the distances between their embeddings naturally leads to our ranking formulation. Taking into account this", "natural ranking from each node's point of view, we learn more powerful embeddings since we incorporate information about the network structure beyond first and second order proximity.Furthermore, when node attributes (e.g. text) are available our method is able to leverage them to easily generate embeddings for previously unseen nodes without additional training. In other words, Graph2Gauss", "is inductive, which is a significant benefit over existing methods that are inherently transductive and do not naturally generalize to unseen nodes. This desirable inductive property", "comes from the fact that we are learning an encoder that maps the nodes' attributes to embeddings.The main contributions of our approach are summarized as follows: a) We embed nodes as Gaussian distributions", "allowing us to capture uncertainty. b) Our unsupervised personalized ranking formulation", "exploits the natural ordering of the nodes capturing the network structure at multiple scales.c) We propose an inductive method that generalizes to unseen nodes and is applicable to different types of graphs: plain/attributed, directed/undirected.", "Inductive learning.", "While during learning we need both the network structure (to evaluate the ranking loss) and the attributes, once the learning concludes, the embedding for a node can be obtained solely based on its attributes.", "This enables our method to easily handle the issue of obtaining representations for new nodes that were not part of the network during training.", "To do so we simply pass the attributes of the new node through our learned deep encoder.", "Most approaches cannot handle this issue at all, with a notable exception being SDNE and GraphSAGE .", "However, both approaches require the edges of the new node to get the node's representation, and cannot handle nodes that have no existing connections.", "In contrast, our method can handle even such nodes, since after the model is learned we rely only on the attribute information.Plain graph embedding.", "Even though attributed graphs are often found in the real-world, sometimes it is desirable to analyze plain graphs.", "As already discussed, our method easily handles plain graphs, when the attributes are not available, by using one-hot encoding of the nodes instead.", "As we later show in the experiments we are able to learn useful representations in this scenario, even outperforming some attributed approaches.", "Naturally, in this case we lose the inductive ability to handle unseen nodes.", "We compare the one-hot encoding version, termed G2G oh, with our full method G2G that utilizes the attributes, as well as all remaining competitors.Encoder architecture.", "Depending on the type of the node attributes (e.g. images, text) we could in principle use CNNs/RNNs to process them.", "We could also easily incorporate any of the proposed graph convolutional layers inheriting their benefits.", "However, we observe that in practice using simple feed-forward architecture with rectifier units is sufficient, while being much faster and easier to train.", "Better yet, we observed that Graph2Gauss is not sensitive to the choice of hyperparameters such as number and size of hidden layers.", "We provide more detailed information and sensible defaults in the appendix.Complexity.", "The time complexity for computing the original loss is O(N 3 ) where N is the number of nodes.", "Using our node-anchored sampling strategy, the complexity of the stochastic version is O(K 2 N ) where K is the maximum distance considered.", "Since a small value of K ≤ 2 consistently showed good performance, K 2 becomes negligible and thus the complexity is O(N ), meaning linear in the number of nodes.", "This coupled with the small number of epochs T needed for convergence (T ≤ 2000 for all shown experiments, see e.g. FIG2 ) and an efficient GPU implementation also made our method faster than most competitors in terms of wall-clock time.", "We proposed Graph2Gauss -the first unsupervised approach that represents nodes in attributed graphs as Gaussian distributions and is therefore able to capture uncertainty.", "Analyzing the uncertainty reveals the latent dimensionality of a graph and gives insight into the neighborhood diversity of a node.", "Since we exploit the attribute information of the nodes we can effortlessly generalize to unseen nodes, enabling inductive reasoning.", "Graph2Gauss leverages the natural ordering of the nodes w.r.t. their neighborhoods via a personalized ranking formulation.", "The strength of the learned embeddings has been demonstrated on several tasks -specifically achieving high link prediction performance even in the case of low dimensional embeddings.", "As future work we aim to study personalized rankings beyond the ones imposed by the shortest path distance." ]

[ 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.21052631735801697, 0.08888888359069824, 0.5777777433395386, 0.052631575614213943, 0.10810810327529907, 0.0952380895614624, 0, 0.14999999105930328, 0.1463414579629898, 0.12903225421905518, 0.05405404791235924, 0.05714285373687744, 0.07999999821186066, 0.17142856121063232, 0.31578946113586426, 0.19354838132858276, 0.0476190410554409, 0.2666666507720947, 0.2380952388048172, 0.23999999463558197, 0.1666666567325592, 0.14999999105930328, 0.0833333283662796, 0.1428571343421936, 0.2978723347187042, 0.3571428656578064, 0.12765957415103912, 0.043478257954120636, 0.10256409645080566, 0, 0.060606054961681366, 0.10256409645080566, 0.04878048226237297, 0.11764705181121826, 0.05128204822540283, 0.10810810327529907, 0.19999998807907104, 0.09999999403953552, 0.10810810327529907, 0.1875, 0.09999999403953552, 0.10526315122842789, 0.13793103396892548, 0.05882352590560913, 0, 0.1395348757505417, 0.035087715834379196, 0.44999998807907104, 0.1818181723356247, 0.11764705181121826, 0.1764705777168274, 0.04999999329447746, 0.05882352590560913 ]

[ " We embed nodes in a graph as Gaussian distributions allowing us to capture uncertainty about their representation." ]

[ "While great progress has been made at making neural networks effective across a wide range of tasks, many are surprisingly vulnerable to small, carefully chosen perturbations of their input, known as adversarial examples.", "In this paper, we advocate for and experimentally investigate the use of logit regularization techniques as an adversarial defense, which can be used in conjunction with other methods for creating adversarial robustness at little to no cost.", "We demonstrate that much of the effectiveness of one recent adversarial defense mechanism can be attributed to logit regularization and show how to improve its defense against both white-box and black-box attacks, in the process creating a stronger black-box attacks against PGD-based models.\n", "Neural networks, despite their high performance on a variety of tasks, can be brittle.", "Given data intentionally chosen to trick them, many deep learning models suffer extremely low performance.", "This type of data, commonly referred to as adversarial examples, represent a security threat to any machine learning system where an attacker has the ability to choose data input to a model, potentially allowing the attacker to control a model's behavior.Today, adversarial examples are typically created by small, but carefully chosen transformations of data that models are otherwise high-performant on.", "This is primarily due to the ease of experimentation with existing datasets BID4 , though the full threat of adversarial examples is only limited by the ability and creativity of an attacker's example generation process.Even with the limited threat models considered in current research, performance on adversarially chosen examples can be dramatically worse than unperturbed data -for example, white-box accuracy on adversarially chosen examples for the CIFAR-10 image classification task BID10 ) is lower than 50%, even for the most robust defenses known today BID12 BID9 , while unperturbed accuracy can be as high as 98.", "5% Cubuk et al. (2018) .Current", "defenses against adversarial examples generally come in one of a few flavors. Perhaps", "the most common approach is to generate adversarial examples as part of the training procedure and explicitly train on them (\"adversarial training\"). Another", "approach is to transform the model's input representation in a way that thwarts an attacker's adversarial example construction mechanism. While these", "methods can be effective, care must be taken to make sure that they are not merely obfuscating gradients BID1 . Last, generative", "models can be built to model the original data distribution, recognizing when the input data is out of sample and potentially correcting it BID18 BID16 . Of these, perhaps", "the most robust today is adversarial logit pairing BID9 , which extends the adversarial training work of BID12 by incorporating an additional term to make the logits (pre-softmax values) of an unperturbed and adversarial example more similar.In this work, we show that adversarial logit pairing derives a large fraction of its benefits from regularizing the model's logits toward zero, which we demonstrate through simple and easy to understand theoretical arguments in addition to empirical demonstration. Investigating this", "phenomenon further, we examine two alternatives for logit regularization, finding that both result in improved robustness to adversarial examples, sometimes surprisingly so -for example, using the right amount of label smoothing BID21 can result in greater than 40% robustness to a projected gradient descent (PGD) attack BID12 on CIFAR-10 while training only on the original, unperturbed training examples, and is also a compelling black-box defense. We then present an", "alternative formulation of adversarial logit pairing that separates the logit pairing and logit regularization effects, improving the defense. The end result of", "these investigations is a defense that sets a new state-of-the-art for PGD-based adversaries on CIFAR-10 for both white box and black box attacks, while requiring little to no computational overhead on top of adversarial training.", "In this work, we have shown the usefulness of logit regularization for improving the robustness of neural networks to adversarial examples.", "We first presented an analysis of adversarial logit pairing, the current state-of-the-art in adversarial defense, showing that roughly half of its improvement over adversarial training can be attributed to a non-obvious logit regularization effect.", "Based on this, we investigated two other forms of logit regularization, demonstrating the benefits of both, and then presented an alternative method for adversarial logit pairing that more cleanly decouples the logit pairing and logit regularization effects while also improving performance.By combining these logit regularization techniques together, we were able to create both a stronger defense against white-box PGD-based attacks and also a stronger attack against PGD-based defenses, both of which come at almost no additional cost to PGD-based adversarial training.", "We also showed the surprising strength of label smoothing as a black-box defense and its corresponding weakness to only highly-optimized white-box attacks.We anticipate that future work will push the limits of logit regularization even further to improve defenses against adversarial examples, possibly using more techniques originally devised for other purposes BID14 .", "We also hope that these investigations will yield insights into training adversarially-robust models without the overhead of multi-step adversarial training, an obstacle that has made it challenge to scale up adversarial defenses to larger datasets without a sizable computational budget." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0 ]

[ 0.08888888359069824, 0.25, 0.23999999463558197, 0.07407406717538834, 0, 0.09677419066429138, 0.11764705926179886, 0, 0.23076923191547394, 0.22857142984867096, 0.11764705181121826, 0.060606054961681366, 0.1538461446762085, 0.10958904027938843, 0.10958904027938843, 0.3333333432674408, 0.13636362552642822, 0.25, 0.1860465109348297, 0.13513512909412384, 0.2295081913471222, 0.16326530277729034 ]

[ "Logit regularization methods help explain and improve state of the art adversarial defenses" ]

[ "In deep learning, performance is strongly affected by the choice of architecture\n", "and hyperparameters.", "While there has been extensive work on automatic hyperpa-\n", "rameter optimization for simple spaces, complex spaces such as the space of deep\n", "architectures remain largely unexplored.", "As a result, the choice of architecture is\n", "done manually by the human expert through a slow trial and error process guided\n", "mainly by intuition.", "In this paper we describe a framework for automatically\n", "designing and training deep models.", "We propose an extensible and modular lan-\n", "guage that allows the human expert to compactly represent complex search spaces\n", "over architectures and their hyperparameters.", "The resulting search spaces are tree-\n", "structured and therefore easy to traverse.", "Models can be automatically compiled to\n", "computational graphs once values for all hyperparameters have been chosen.", "We\n", "can leverage the structure of the search space to introduce different model search\n", "algorithms, such as random search, Monte Carlo tree search (MCTS), and sequen-\n", "tial model-based optimization (SMBO).", "We present experiments comparing the\n", "different algorithms on CIFAR-10 and show that MCTS and SMBO outperform\n", "random search.", "We also present experiments on MNIST, showing that the same\n", "search space achieves near state-of-the-art performance with a few samples.", "These\n", "experiments show that our framework can be used effectively for model discov-\n", "ery, as it is possible to describe expressive search spaces and discover competitive\n", "models without much effort from the human expert.", "Code for our framework and\n", "experiments has been made publicly available", "Deep learning has seen a surge in popularity due to breakthroughs in applications such as computer vision, natural language processing, and reinforcement learning BID12 Karpathy & FeiFei, 2015; BID24 ).", "An important observation in much of the recent work is that complex architectures are important for achieving high performance BID12 BID20 .", "Larger datasets and more powerful computing infrastructures are likely to increase our ability to effectively train larger, deeper, and more complex architectures.", "However, improving the performance of a neural network is not as simple as adding more layers or parameters-it often requires clever ideas such as creating more branches or adding skip connections BID12 .", "Even popular techniques such as dropout BID27 and batch normalization BID14 do not always lead to better performance, and need to be judiciously applied to be helpful.Currently, choosing appropriate values for these architectural hyperparameters requires close supervision by a human expert, in a trial and error manual search process largely guided by intuition.", "The expert is burdened by having to make the large number of choices involved in the specification of a deep model.", "Choices interact in non-obvious ways and strongly impact performance.", "The typical workflow has the expert specify a single model, train it, and compute a validation score.", "Based on the validation score, previous experience, and information gathered during training, the expert decides if the trained model is satisfactory or not.", "If the model is considered unsatisfactory, the expert has to think about model variations that may lead to better performance.From the perspective of the expert, it would be convenient to search over architectures automatically, just as we search over simple scalar hyperparameters, such as the learning rate and the regularization coefficient.", "Ideally, the expert would have control in setting up the search space to incorporate inductive biases about the task being solved and constraints about computational resources.", "Prior to this work, achieving this goal was hard because expressing model search spaces using general hyperparameter optimization tools requires the human expert to manually distill a set of relevant scalar architectural hyperparameters.The main contributions of our work are 1. a modular, compositional, and extensible language for compactly representing expressive search spaces over models that", "(a) gives control to the human expert over what model variations to consider;", "(b) makes it easy to automatically search for performant models in the search space;", "(c) allows models to be directly compiled to computational graphs without the human expert having to write additional code.", "2. model search algorithms that rely on the tree-structured search spaces induced by our language to systematically and efficiently search for performant models; namely, we", "(a) show that by using constructs in our language, even random search can be effective;", "(b) compare different model search algorithms experimentally, and show that random search is outperformed by algorithms that leverage the structure of the search space to generalize more effectively across different models.The main differences between our work and previous work are that we develop a modular, composable and extensible language, focusing on the problem of searching over deep architectures.", "This focus allows the expert to compactly set up a search space, search over it, and automatically compile models to their corresponding computational graphs.", "Our language can be seen as an effort to combine the functionalities of a deep model specification language (e.g., Tensorflow BID0 ) and a structured hyperparameter search language (e.g., Hyperopt BID32 ).", "We described a framework for automatically designing and training deep models.", "This framework consists of three fundamental components: the model search space specification language, the model search algorithm, and the model evaluation algorithm.", "The model search space specification language is composable, modular, and extensible, and allows us to easily define expressive search spaces over architectures.", "The model evaluation algorithm determines how to compute a score for a model in the search space.", "Models can be automatically compiled to their corresponding computational graphs.", "Using the model search space specification language and the model evaluation algorithm, we can introduce model search algorithms for exploring the search space.", "Using our framework, it is possible to do random search over interesting spaces of architectures without much effort from the expert.", "We also described more complex model search algorithms, such as MCTS, MCTS with tree restructuring, and SMBO.", "We present experiments on CIFAR-10 comparing different model search algorithms and show that MCTS with tree restructuring and SMBO outperform random search.", "Code for our framework and experiments has been made publicly available.", "We hope that this paper will lead to more work and better tools for automatic architecture search." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0, 0, 0.1764705777168274, 0.07999999821186066, 0.06896550953388214, 0.11428570747375488, 0, 0.20000000298023224, 0.07692307233810425, 0.2142857164144516, 0.1818181723356247, 0.23076923191547394, 0.14814814925193787, 0.14814814925193787, 0.07407406717538834, 0.06451612710952759, 0.1875, 0.12121211737394333, 0, 0.07692307233810425, 0.12903225421905518, 0.06451612710952759, 0.12903225421905518, 0.12121211737394333, 0.3529411852359772, 0, 0.1538461446762085, 0, 0.16326530277729034, 0.09756097197532654, 0.14999999105930328, 0.0833333283662796, 0.20588235557079315, 0.14999999105930328, 0.06666666269302368, 0.10810810327529907, 0.0952380895614624, 0.22580644488334656, 0.13636362552642822, 0.28169015049934387, 0.1818181723356247, 0.1764705777168274, 0.052631575614213943, 0.3636363446712494, 0.0555555522441864, 0.2647058665752411, 0.23255813121795654, 0.23999999463558197, 0.25, 0.15789473056793213, 0.4390243887901306, 0.277777761220932, 0.06451612710952759, 0.3243243098258972, 0.2380952388048172, 0.21052631735801697, 0.24390242993831635, 0.1249999925494194, 0.2631579041481018 ]

[ "We describe a modular and composable language for describing expressive search spaces over architectures and simple model search algorithms applied to these search spaces. " ]

[ "Deep learning networks have achieved state-of-the-art accuracies on computer vision workloads like image classification and object detection.", "The performant systems, however, typically involve big models with numerous parameters.", "Once trained, a challenging aspect for such top performing models is deployment on resource constrained inference systems -- the models (often deep networks or wide networks or both) are compute and memory intensive.", "Low precision numerics and model compression using knowledge distillation are popular techniques to lower both the compute requirements and memory footprint of these deployed models.", "In this paper, we study the combination of these two techniques and show that the performance of low precision networks can be significantly improved by using knowledge distillation techniques.", "We call our approach Apprentice and show state-of-the-art accuracies using ternary precision and 4-bit precision for many variants of ResNet architecture on ImageNet dataset.", "We study three schemes in which one can apply knowledge distillation techniques to various stages of the train-and-deploy pipeline.", "Background: Today's high performing deep neural networks (DNNs) for computer vision applications comprise of multiple layers and involve numerous parameters.", "These networks have O(Giga-FLOPS) compute requirements and generate models which are O(Mega-Bytes) in storage BID4 .", "Further, the memory and compute requirements during training and inference are quite different BID23 .", "Training is performed on big datasets with large batch-sizes where memory footprint of activations dominates the model memory footprint.", "On the other hand, batchsize during inference is typically small and the model's memory footprint dominates the runtime memory requirements.Because of complexity in compute, memory and storage requirements, training phase of the networks is performed on CPU and/or GPU clusters in a distributed computing environment.", "Once trained, a challenging aspect is deployment of trained models on resource constrained inference systems such as portable devices or sensor networks, and for applications in which real-time predictions are required.", "Performing inference on edge-devices comes with severe constraints on memory, compute and power.", "Additionally, ensemble based methods, which one can potentially use to get improved accuracy predictions, become prohibitive in resource constrained systems.Quantization using low-precision numerics BID37 BID45 BID21 BID24 BID10 BID46 BID26 BID35 BID23 and model compression BID3 BID16 BID27 have emerged as popular solutions for resource constrained deployment scenarios.", "With quantization, a low-precision version of network model is generated and deployed on the device.", "Operating in lower precision mode reduces compute as well as data movement and storage requirements.", "However, majority of existing works in low-precision DNNs sacrifice accuracy over baseline full-precision networks.", "With model compression, a smallerIn the second scheme, we start with a full-precision trained network and transfer knowledge from this trained network continuously to train a low-precision network from scratch.", "We find that the low-precision network converges faster (albeit to similar accuracies as the first scheme) when a trained complex network guides its training.In the third scheme, we start with a trained full-precision large network and an apprentice network that has been initialised with full-precision weights.", "The apprentice network's precision is lowered and is fine-tuned using knowledge distillation techniques.", "We find that the low-precision network's accuracy marginally improves and surpasses the accuracy obtained via the first scheme.", "This scheme then sets the new state-of-the-art accuracies for the ResNet models at ternary and 4-bit precision.Overall, the contributions of this paper are the techniques to obtain low-precision DNNs using knowledge distillation technique.", "Each of our scheme produces a low-precision model that surpasses the accuracy of the equivalent low-precision model published to date.", "One of our schemes also helps a low-precision model converge faster.", "We envision these accurate low-precision models to simplify the inference deployment process on resource constrained systems and even otherwise on cloud-based deployment systems.", "In scheme-A, we use a teacher network that is always as large or larger in number of parameters than the student network.", "We experimented with a ternary ResNet-34 student network which was paired with a full-precision ResNet-18.", "The ternary model for ResNet-34 is about 8.5x smaller in size compared to the full-precision ResNet-18 model.", "The final trained accuracy of the ResNet-34 ternary model with this setup is 2.7% worse than that obtained by pairing the ternary ResNet-34 network with a ResNet-50 teacher network.", "This suggests that the distillation scheme works only when the teacher network is higher in accuracy than the student network (and not necessarily bigger in capacity).", "Further, the benefit from using a larger teacher network saturates at some point.", "This can be seen by picking up a precision point, say \"32A, 2W\" and looking at the error rates along the row in TAB2 , 2 and 3.One concern, we had in the early stages of our investigation, with joint training of a low-precision small network and a high precision large network was the influence of the small network's accuracy on the accuracy of the large network.", "When using the joint cost function, the smaller network's probability scores are matched with the predictions from the teacher network.", "The joint cost is added as a term to the total loss function (equation 1).", "This led us to posit that the larger network's learning capability will be affected by the inherent impairment in the smaller low-precision network.", "Further, since the smaller student network learns form the larger teacher network, a vicious cycle might form where the student network's accuracy will further drop because the teacher network's learning capability is being impeded.", "However, in practice, we did not see this phenomenon occurring -in each case where the teacher network was jointly trained with a student network, the accuracy of the teacher network was always within 0.1% to 0.2% of the accuracy of the teacher network without it jointly supervising a student network.", "This could be because of our choice of α, β and γ values.In Section 4, we mentioned about temperature, τ , for Softmax function and hyper-parameters α = 1, β = 0.5 and γ = 0.5.", "Since, we train directly on the logits of the teacher network, we did not have to experiment with the appropriate value of τ .", "τ is required when training on the soft targets produced by the teacher network.", "Although we did not do extensive studies experimenting with training on soft targets as opposed to logits, we find that τ = 1 gives us best results when training on soft targets.", "BID16 mention that when the student network is significantly smaller than the teacher network, small values of τ are more effective than large values.", "For few of the low-precision configurations, we experimented with α = β = γ = 1, and, α = 0.9, β = 1 and γ = 0.1 or 0.3.", "Each of these configurations, yielded a lower performance model compared to our original choice for these parameters.For the third term in equation 1, we experimented with a mean-squared error loss function and also a loss function with logits from both the student and the teacher network (i.e. H(z T , z A )).", "We did not find any improvement in accuracy compared to our original choice of the cost function formulation.", "A thorough investigation of the behavior of the networks with other values of hyper-parameters and different loss functions is an agenda for our future work.Overall, we find the distillation process to be quite effective in getting us high accuracy low-precision models.", "All our low-precision models surpass previously reported low-precision accuracy figures.", "For example, TTQ scheme achieves 33.4% Top-1 error rate for ResNet-18 with 2-bits weight.", "Our best ResNet-18 model, using scheme-A, with 2-bits weight achieves ∼31.5% error rate, improving the model accuracy by ∼2% over TTQ.", "Similarly, the scheme in BID22 achieves 29.2% Top-1 error with 2-bits weight and 8-bits activation.", "The best performing Apprentice network at this precision achieves 27.2% Top-1 error.", "For Scheme-B and Scheme-C, which we describe next, Scheme-A serves as the new baseline.", "As mentioned earlier, low-precision is a form of model compression.", "There are many works which target network sparsification and pruning techniques to compress a model.", "With ternary precision models, the model size reduces by a factor of 2/32 compared to full-precision models.", "With Apprentice, we show how one can get a performant model with ternary precision.", "Many works targeting network pruning and sparsification target a full-precision model to implement their scheme.To be comparable in model size to ternary networks, a full-precision model needs to be sparsified by 93.75%.", "Further, to be effective, a sparse model needs to store a key for every non-zero value denoting the position of the value in the weight tensor.", "This adds storage overhead and a sparse model needs to be about 95% sparse to be at-par in memory size as a 2-bit model.", "Note that ternary precision also has inherent sparsity (zero is a term in the ternary symbol dictionary) -we find our ternary models to be about 50% sparse.", "In work by and BID12 , sparsification of full-precision networks is proposed but the sparsity achieved is less than 93.75%.", "Further, the network accuracy using techniques in both these works lead to larger degradation in accuracy compared to our ternary models.", "Overall, we believe, our ternary precision models to be state-of-the-art not only in accuracy (we better the accuracy compared to prior ternary precision models) but also when one considers the size of the model at the accuracy level achieved by low-precision or sparse networks.", "While low-precision networks have system-level benefits, the drawback of such models is degraded accuracy when compared to full-precision models.", "We present three schemes based on knowledge distillation concept to improve the accuracy of low-precision networks and close the gap between the accuracy of these models and full-precision models.", "Each of the three schemes improve the accuracy of the low-precision network configuration compared to prior proposals.", "We motivate the need for a smaller model size in low batch, real-time and resource constrained inference deployment systems.", "We envision the low-precision models produced by our schemes to simplify the inference deployment process on resource constrained systems and on cloud-based deployment systems where low latency is a critical requirement." ]

[ 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.1538461446762085, 0, 0.1538461446762085, 0.260869562625885, 0.4583333432674408, 0.3636363446712494, 0.2926829159259796, 0.1904761791229248, 0.10810810327529907, 0.11428570747375488, 0.10256409645080566, 0.1355932205915451, 0.11320754140615463, 0.05882352590560913, 0.11594202369451523, 0.1621621549129486, 0.1111111044883728, 0.1666666567325592, 0.17391303181648254, 0.1355932205915451, 0.23529411852359772, 0.2702702581882477, 0.37735849618911743, 0.21052631735801697, 0.060606054961681366, 0.1428571343421936, 0.1395348757505417, 0.11428570747375488, 0.1538461446762085, 0.21276594698429108, 0.13636362552642822, 0.05714285373687744, 0.17142856121063232, 0.05128204822540283, 0.05405404791235924, 0.09302324801683426, 0.08163265138864517, 0.10526315122842789, 0.11538460850715637, 0.09756097197532654, 0.05714285373687744, 0.04081632196903229, 0.1395348757505417, 0.1395348757505417, 0.11764705181121826, 0.19999998807907104, 0.19999998807907104, 0.06451612710952759, 0.05405404791235924, 0.09090908616781235, 0.10526315122842789, 0.05714285373687744, 0.1666666567325592, 0.0624999962747097, 0.10810810327529907, 0.20512819290161133, 0.2222222238779068, 0.08163265138864517, 0.1395348757505417, 0.04878048226237297, 0.1702127605676651, 0.1904761791229248, 0.19999998807907104, 0.24137930572032928, 0.19999998807907104, 0.35555556416511536, 0.2222222238779068, 0.24390242993831635, 0.16326530277729034 ]

[ "We show that knowledge transfer techniques can improve the accuracy of low precision networks and set new state-of-the-art accuracy for ternary and 4-bits precision. " ]

[ "Deep neural networks (DNN) are widely used in many applications.", "However, their deployment on edge devices has been difficult because they are resource hungry.", "Binary neural networks (BNN) help to alleviate the prohibitive resource requirements of DNN, where both activations and weights are limited to 1-bit.", "We propose an improved binary training method (BNN+), by introducing a regularization function that encourages training weights around binary values.", "In addition to this, to enhance model performance we add trainable scaling factors to our regularization functions.", "Furthermore, we use an improved approximation of the derivative of the sign activation function in the backward computation.", "These additions are based on linear operations that are easily implementable into the binary training framework.", "We show experimental results on CIFAR-10 obtaining an accuracy of 86.5%, on AlexNet and 91.3% with VGG network.", "On ImageNet, our method also outperforms the traditional BNN method and XNOR-net, using AlexNet by a margin of 4% and 2% top-1 accuracy respectively.", "Deep neural networks (DNNs) have demonstrated success for many supervised learning tasks ranging from voice recognition to object detection BID26 BID11 .", "The focus has been on increasing accuracy, in particular for image tasks, where deep convolutional neural networks (CNNs) are widely used.", "However, their increasing complexity poses a new challenge, and has become an impediment to widespread deployment in many applications; specifically when trying to deploy such models to resource constrained and lower-power devices.", "A typical DNN architecture contains tens to thousands of layers, resulting in millions of parameters.", "As an example, AlexNet BID16 requires 200MB of memory, VGGNet BID26 requires 500MB memory.", "Large model sizes are further exacerbated by their computational cost, requiring GPU implementation to allow real-time inference.", "Such requirements evidently cannot be accustomed by edge devices as they have limited memory, computation power, and battery.", "This motivated the community to investigate methods for compressing and reducing computation cost of DNNs.To make DNNs compatible with the resource constraints of low power devices, there have been several approaches developed, such as network pruning BID17 , architecture design BID25 , and quantization BID0 BID4 .", "In particular, weight compression using quantization can achieve very large savings in memory, where binary (1-bit), and ternary (2-bit) approaches have been shown to obtain competitive accuracy BID10 BID31 BID29 .", "Using such schemes reduces model sizes by 8x to 32x depending on the bit resolution used for computations.", "In addition to this, the speed by quantizing the activation layers.", "In this way, both the weights and activations are quantized so that one can replace the expensive dot products and activation function evaluations with bitwise operations.", "This reduction in bit-width benefits hardware accelerators such as FPGAs and neural network chips.An issue with using low-bit DNNs is the drastic drop in accuracy compared to its full precision counterpart, and this is made even more severe upon quantizing the activations.", "This problem is largely due to noise and lack of precision in the training objective of the networks during back-propagation BID19 .", "Although, quantizing the weights and activations have been attracting large interests thanks to their computational benefits, closing the gap in accuracy between the full precision and the quantized version remains a challenge.", "Indeed, quantizing weights cause drastic information loss and make neural networks harder to train due to a large number of sign fluctuations in the weights.", "Therefore, how to control the stability of this training procedure is of high importance.", "In theory, it is infeasible to back-propagate in a quantized setting as the weights and activations employed are discontinuous and discrete.", "Instead, heuristics and approximations are proposed to match the forward and backward passes.", "Often weights at different layers of DNNs follow a certain structure.", "How to quantize the weights locally, and maintaining a global structure to minimize a common cost function is important BID18 .Our", "contribution consists of three ideas that can be easily implemented in the binary training framework presented by BID10 to improve convergence and generalization accuracy of binary networks. First", ", the activation function is modified to better approximate the sign function in the backward pass, second we propose two regularization functions that encourage training weights around binary values, and lastly a scaling factor is introduced in both the regularization term as well as network building blocks to mitigate accuracy drop due to hard binarization. Our", "method is evaluated on CIFAR-10 and ImageNet datasets and compared to other binary methods. We", "show accuracy gains to traditional binary training.", "We proposed two regularization terms (4) and (5) and an activation term (2) with a trainable parameter β.", "We run several experiments to better understand the effect of the different modifications to the training method, especially using different regularization and asymptote parameters β.", "The parameter β is trainable and would add one equation through back-propagation.", "However, we fixed β throughout our experiments to explicit values.", "The results are summarized in TAB1 .Through", "our experiments, we found that adding regularizing term with heavy penalization degrades the networks ability to converge, as the term would result in total loss be largely due to the regu- larizing term and not the target cross entropy loss. Similarly", ", the regularizing term was set to small values in BID29 . As a result", ", we set λ with a reasonably small value 10 −5 − 10 −7 , so that the scales move slowly as the weights gradually converge to stable values. Some preliminary", "experimentation was to gradually increase the regularization with respect to batch iterations updates done in training, though this approach requires careful tuning and was not pursued further.From TAB1 , and referring to networks without regularization, we see the benefit of using SwishSign approximation versus the STE. This was also noted", "in , where their second approximation provided better results. There is not much difference", "between using R 1 versus R 2 towards model generalization although since the loss metric used was the cross-entropy loss, the order of R 1 better matches the loss metric. Lastly, it seems moderate values", "of β is better than small or large values. Intuitively, this happens because", "for small values of β, the gradient approximation is not good enough and as β increases the gradients become too large, hence small noise could cause large fluctuations in the sign of the weights.We did not compare our network with that of as they introduce a shortcut connection that proves to help even the full precision network. As a final remark, we note that the", "learning rate is of great importance and properly tuning this is required to achieve convergence. Table 3 summarizes the best results", "of the ablation study and compares with BinaryNet, XNOR-Net, and ABC-Net. Table 3 : Comparison of top-1 and top-5", "accuracies of our method BNN+ with BinaryNet, XNORNet and ABC-Net on ImageNet, summarized from TAB1 . The results of BNN, XNOR, & ABC-Net are", "reported from the corresponding papers BID23 BID10 BID29 . Results for ABC-NET on AlexNet were not", "available, and so is not reported.", "To summarize we propose three incremental ideas that help binary training:", "i) adding a regularizer to the objective function of the network,", "ii) trainable scale factors that are embedded in the regularizing term and", "iii) an improved approximation to the derivative of the sign activation function.", "We obtain competitive results by training AlexNet and Resnet-18 on the ImageNet dataset.", "For future work, we plan on extending these to efficient models such as CondenseNet BID9 , MobileNets BID8 , MnasNet BID28 and on object recognition tasks." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.05882352590560913, 0, 0.08888888359069824, 0.2380952388048172, 0, 0.1538461446762085, 0.20512820780277252, 0.1860465109348297, 0.08695651590824127, 0.08888888359069824, 0.13333332538604736, 0.037735845893621445, 0, 0.05405404791235924, 0, 0, 0.09090908616781235, 0.07407406717538834, 0.0952380895614624, 0.05882352590560913, 0.1249999925494194, 0.1904761791229248, 0.1860465109348297, 0.19230768084526062, 0.08510638028383255, 0.10810810327529907, 0.045454539358615875, 0.0555555522441864, 0, 0.04651162400841713, 0.23999999463558197, 0.17142856121063232, 0.052631575614213943, 0.19354838132858276, 0.09756097197532654, 0.08888888359069824, 0.0555555522441864, 0, 0.06451612710952759, 0.13793103396892548, 0.052631575614213943, 0.11538460850715637, 0.0882352888584137, 0, 0.039215680211782455, 0, 0.15789473056793213, 0.045454539358615875, 0.09999999403953552, 0.08888888359069824, 0.10256409645080566, 0, 0.11428570747375488, 0.05882352590560913, 0.1111111044883728, 0.17142856121063232, 0.10810810327529907, 0 ]

[ "The paper presents an improved training mechanism for obtaining binary networks with smaller accuracy drop that helps close the gap with it's full precision counterpart" ]

[ "Clustering is a fundamental machine learning method.", "The quality of its results is dependent on the data distribution.", "For this reason, deep neural networks can be used for learning better representations of the data.", "In this paper, we propose a systematic taxonomy for clustering with deep learning, in addition to a review of methods from the field.", "Based on our taxonomy, creating new methods is more straightforward.", "We also propose a new approach which is built on the taxonomy and surpasses some of the limitations of some previous work.", "Our experimental evaluation on image datasets shows that the method approaches state-of-the-art clustering quality, and performs better in some cases.", "Clustering is one of the most fundamental unsupervised machine learning problems.", "Its main goal is to separate data into clusters of similar data points.", "Besides having its own applications, it is beneficial for multiple other fundamental tasks.", "For instance, it can serve for automatic data labeling for supervised learning and as a pre-processing step for data visualization and analysis.However, the performance of clustering algorithms is dependent on the type of the input data, such that different problems and datasets could require different similarity measures and different separation techniques.", "As a result, dimensionality reduction and representation learning have been extensively used alongside clustering, in order to map the input data into a feature space where separation is easier with respect to the problem's context.", "Using deep neural networks (DNNs), it is possible to learn non-linear mappings allowing to transform the data into more clustering-friendly representations.In the past, dimensionality reduction (or representation learning) and clustering have been treated separately, and sequentially applied on the data BID3 BID22 BID23 .", "However, recent research has shown that jointly optimizing for both problems can achieve decent results BID20 BID28 BID29 BID13 .One", "of our main contributions is the formulation of a taxonomy of methods that use deep learning for clustering. Our", "taxonomy facilitates the overview of existing methods and the creation of new ones by using the best properties of the existing ones in a modular manner.Based on the taxonomy, we propose a new method that combines advantageous properties of some existing methods. We", "use an autoencoder-based method for learning better representations of the data which are clustering-friendly, with a state-of-the-art training procedure. The", "training has two phases, the first one being standard autoencoder training with the mean squared error reconstruction loss, and the second one is based on a loss function combining the reconstruction loss and a clustering-specific loss. Moreover", ", in the second phase, we alternate between optimizing the network model, and updating the clustering assignments.The rest of the paper is organized as follows: the taxonomy of clustering with deep learning and the corresponding building blocks is described in Section 2. In Section", "3, several related methods are briefly described and compared based on the taxonomy. Subsequently", ", in Section 4, a new method is proposed and discussed based on the building blocks of the taxonomy. Results of", "the proposed method are shown in Section 5, followed by conclusions in Section 6.", "In this work, we present a taxonomy for clustering with deep learning, identifying the general framework, and discussing different building blocks and possible options.", "In addition, a summary of methods in the field and their specific use of the taxonomy is presented alongside a general comparison of many of these methods.", "Using this taxonomy and the summary of previous methods, generating new methods is clearer and easier and can be done by creating new combinations of the taxonomy's building blocks.", "Moreover, we present a new method to the field, which is based on such a new combination.", "Our method overcomes the limitations of several previous ones, approaches state-ofthe-art performance and performs better in some cases." ]

[ 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0, 0, 0.23999999463558197, 0.19354838132858276, 0, 0, 0.06896550953388214, 0, 0.0952380895614624, 0, 0.03999999538064003, 0.04878048598766327, 0.2083333283662796, 0, 0.1538461446762085, 0.05128204822540283, 0, 0, 0.09302325546741486, 0, 0, 0, 0.1249999925494194, 0, 0, 0.0833333283662796, 0 ]

[ "Unifying framework to perform clustering using deep neural networks" ]

[ "Generative models often use human evaluations to determine and justify progress.", "Unfortunately, existing human evaluation methods are ad-hoc: there is currently no standardized, validated evaluation that: (1) measures perceptual fidelity, (2) is reliable, (3) separates models into clear rank order, and (4) ensures high-quality measurement without intractable cost.", "In response, we construct Human-eYe Perceptual Evaluation (HYPE), a human metric that is (1) grounded in psychophysics research in perception, (2) reliable across different sets of randomly sampled outputs from a model, (3) results in separable model performances, and (4) efficient in cost and time.", "We introduce two methods.", "The first, HYPE-Time, measures visual perception under adaptive time constraints to determine the minimum length of time (e.g., 250ms) that model output such as a generated face needs to be visible for people to distinguish it as real or fake.", "The second, HYPE-Infinity, measures human error rate on fake and real images with no time constraints, maintaining stability and drastically reducing time and cost.", "We test HYPE across four state-of-the-art generative adversarial networks (GANs) on unconditional image generation using two datasets, the popular CelebA and the newer higher-resolution FFHQ, and two sampling techniques of model outputs.", "By simulating HYPE's evaluation multiple times, we demonstrate consistent ranking of different models, identifying StyleGAN with truncation trick sampling (27.6% HYPE-Infinity deception rate, with roughly one quarter of images being misclassified by humans) as superior to StyleGAN without truncation (19.0%) on FFHQ.", "Historically, likelihood-based estimation techniques served as the de-facto evaluation metric for generative models BID18 BID5 .", "But recently, with the application of generative models to complex tasks such as image and text generation BID14 BID34 , likelihood or density estimation grew no longer tractable BID46 .", "Moreover, for high-dimensional problems, even likelihood-based evaluation has been called into question BID46 .", "Consequently, most generative tasks today resort to analyzing model outputs BID41 BID43 BID11 BID21 BID7 BID37 .", "These output evaluation metrics consist of either automatic algorithms that do not reach the ideals of likelihood-based estimation, or ad-hoc human-derived methods that are unreliable and inconsistent BID41 BID11 .Consider", "the well-examined and popular computer vision task of realistic face generation BID14 . Automatic", "algorithms used for this task include Inception Score (IS) BID43 and Fréchet Inception Distance (FID) BID17 . Both have", "been discredited for evaluation on non-ImageNet datasets such as faces BID2 BID40 BID6 BID38 . They are", "also much more sensitive to visual corruptions such as salt and pepper noise than to semantic distortions such as swirled images BID17 . So, while", "automatic metrics are consistent and standardized, they cannot fully capture the semantic side of perceptual fidelity BID6 .Realizing", "the constraints of the available automatic metrics, many generative modeling challenges resort to summative assessments that are completely human BID41 BID43 BID11 . These human", "measures are (1) ad-hoc, each executed in idiosyncrasy without proof of reliability or grounding to theory, and (2) high variance in their estimates BID43 BID11 BID33 . These characteristics", "combine to a lack of reliability, and downstream, (3) a lack of clear separability between models. Theoretically, given", "sufficiently large sample sizes of human evaluators and model outputs, the law of large numbers would smooth out the variance and reach eventual convergence; but this would occur at (4) a high cost and a long delay.In this paper, we present HYPE (HUMAN EYE PERCEPTUAL EVALUATION) that addresses these criteria in turn. It: (1) measures the", "perceptual fidelity of generative model outputs via a grounded method inspired by psychophysics methods in perceptual psychology, (2) is a reliable and consistent estimator, (3) is statistically separable to enable a comparative ranking, and (4) ensures a cost and time efficient method through modern crowdsourcing techniques such as training and aggregation. We present two methods", "of evaluation. The first, called HYPE", "time , is drawn directly from psychophysics literature BID22 ) and displays images using adaptive time constraints to determine the time-limited perceptual threshold a person needs to distinguish real from fake BID9 . The HYPE time score is", "understood as the minimum time, in milliseconds, that a person needs to see the model's output before they can distinguish it as real or fake. Small HYPE time scores", "indicate that model outputs can be identified even at a glance; large scores suggest that people need to dedicate substantial time and attention. The second method, called", "HYPE ∞ , is derived from the first to make it simpler, faster, and cheaper while maintaining reliability. It measures human deception", "from fake images with no time constraints. The HYPE ∞ score is interpretable", "as the rate at which people mistake fake images and real images, given unlimited time to make their decisions.We demonstrate HYPE's performance on unconditional generation of human faces using generative adversarial networks (GANs) BID14 . We evaluate four state-of-the-art", "GANs: WGAN-GP BID16 , BEGAN BID4 , ProGAN BID20 , and the most recent StyleGAN BID21 . First, we track progress across the", "years on the popular CelebA dataset BID28 . We derive a ranking based on perception", "(HYPE time , in milliseconds) and error rate (HYPE ∞ , as a percentage) as follows: StyleGAN (439.4ms, 50.7%), ProGAN (363.7ms, 40.3%), BEGAN (111.1ms, 10.0%), WGAN-GP (100.0ms, 3.8%). A score of 500ms on HYPE time indicates", "that outputs from the model become indistinguishable from real, when shown for 500ms or less, but any more would start to reveal notable differences. A score of 50% on HYPE ∞ represents indistinguishable", "results from real, conditioned on the real training set, while a score above 50% through 100% represents hyper-realism in which generated images appear more real than real ones when drawn from a mixed pool of both. Next, we test StyleGAN trained on the newer FFHQ dataset", "BID21 , comparing between outputs generated when sampled with and without the truncation trick, a technique used to prune low-fidelity generated images BID7 BID21 . We find that outputs generated with the truncation trick", "(363.2ms, 27.6%) significantly outperforms those without it (240.7ms, 19.0%), which runs counter to scores reported by FID.HYPE indicates that GANs have clear, measurable perceptual differences between them. HYPE produces identical rankings between HYPE time and HYPE", "∞ . We also find that even the best eval- Images on the right exhibit", "the highest HYPE scores, the highest human perceptual fidelity. uated model, StyleGAN trained on FFHQ and sampled with the truncation", "trick, only performs at 27.6% HYPE ∞ , suggesting substantial opportunity for improvement. Finally, we show that we can reliably reproduce these results with 95", "% confidence intervals using 30 human evaluators at $60 in a task that takes 10 minutes. While important measures, we do not focus on diversity, overfitting,", "entanglement, training stability, and computational and sample efficiency of the model BID6 BID29 and instead aim to construct the gold standard for human perceptual fidelity.We deploy HYPE as a rapid solution for researchers to measure their generative models, requiring just a single click to produce reliable scores and measure progress. We deploy HYPE at https://hype.stanford.edu, where researchers can upload", "a model and retrieve a HYPE score in 10 minutes for $60. Future work would extend HYPE to adapt to other generative tasks such as", "text generation or abstractive summarization." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.06896550953388214, 0.11320754140615463, 0.20689654350280762, 0, 0.1071428507566452, 0.10256409645080566, 0.1702127605676651, 0.10344827175140381, 0.24242423474788666, 0.12765957415103912, 0.12903225421905518, 0.05882352590560913, 0.043478257954120636, 0.12903225421905518, 0.05714285373687744, 0.11764705181121826, 0, 0, 0.09999999403953552, 0, 0.060606054961681366, 0.09090908616781235, 0.1269841194152832, 0.1666666567325592, 0.11999999731779099, 0.08888888359069824, 0.045454539358615875, 0.14999999105930328, 0.19354838132858276, 0.1071428507566452, 0.05405404791235924, 0.06451612710952759, 0.06896550953388214, 0.08163265138864517, 0.03448275476694107, 0.09090908616781235, 0.072727270424366, 0, 0.17142856121063232, 0.1395348757505417, 0.09090908616781235, 0.20895521342754364, 0.19999998807907104, 0.08695651590824127 ]

[ "HYPE is a reliable human evaluation metric for scoring generative models, starting with human face generation across 4 GANs." ]

[ "Machine translation has recently achieved impressive performance thanks to recent advances in deep learning and the availability of large-scale parallel corpora.", "There have been numerous attempts to extend these successes to low-resource language pairs, yet requiring tens of thousands of parallel sentences.", "In this work, we take this research direction to the extreme and investigate whether it is possible to learn to translate even without any parallel data.", "We propose a model that takes sentences from monolingual corpora in two different languages and maps them into the same latent space.", "By learning to reconstruct in both languages from this shared feature space, the model effectively learns to translate without using any labeled data.", "We demonstrate our model on two widely used datasets and two language pairs, reporting BLEU scores of 32.8 and 15.1 on the Multi30k and WMT English-French datasets, without using even a single parallel sentence at training time.", "Thanks to recent advances in deep learning BID33 BID0 and the availability of large-scale parallel corpora, machine translation has now reached impressive performance on several language pairs .", "However, these models work very well only when provided with massive amounts of parallel data, in the order of millions of parallel sentences.", "Unfortunately, parallel corpora are costly to build as they require specialized expertise, and are often nonexistent for low-resource languages.", "Conversely, monolingual data is much easier to find, and many languages with limited parallel data still possess significant amounts of monolingual data.There have been several attempts at leveraging monolingual data to improve the quality of machine translation systems in a semi-supervised setting BID25 BID15 BID16 BID39 .", "Most notably, BID30 proposed a very effective data-augmentation scheme, dubbed \"back-translation\", whereby an auxiliary translation system from the target language to the source language is first trained on the available parallel data, and then used to produce translations from a large monolingual corpus on the target side.", "The pairs composed of these translations with their corresponding ground truth targets are then used as additional training data for the original translation system.Another way to leverage monolingual data on the target side is to augment the decoder with a language model BID11 .", "And finally, BID3 ; have proposed to add an auxiliary auto-encoding task on monolingual data, which ensures that a translated sentence can be translated back to the original one.", "All these works still rely on several tens of thousands parallel sentences, however.Previous work on zero-resource machine translation has also relied on labeled information, not from the language pair of interest but from other related language pairs BID7 BID17 BID2 or from other modalities BID26 BID22 .", "The only exception is the work by BID29 ; BID28 , where the machine translation problem is reduced to a deciphering problem.", "Unfortunately, their method is limited to rather short sentences and it has only been demonstrated on a very simplistic setting comprising of the most frequent short sentences, or very closely related languages.", "Left (autoencoding) : the model is trained to reconstruct a sentence from a noisy version of it.", "x is the target, C(x) is the noisy input,x is the reconstruction.", "Right (translation): the model is trained to translate a sentence in the other domain.", "The input is a noisy translation (in this case, from source-to-target) produced by the model itself, M , at the previous iteration (t), y = M (t) (x).", "The model is symmetric, and we repeat the same process in the other language.", "See text for more details.In this paper, we investigate whether it is possible to train a general machine translation system without any form of supervision whatsoever.", "The only assumption we make is that there exists a monolingual corpus on each language.", "This set up is interesting for a twofold reason.", "First, this is applicable whenever we encounter a new language pair for which we have no annotation.", "Second, it provides a strong lower bound performance on what any good semi-supervised approach is expected to yield.The key idea is to build a common latent space between the two languages (or domains) and to learn to translate by reconstructing in both domains according to two principles:", "(i) the model has to be able to reconstruct a sentence in a given language from a noisy version of it, as in standard denoising auto-encoders BID36 .(ii", ") The model also learns to reconstruct any source sentence given a noisy translation of the same sentence in the target domain, and vice versa. For", "(ii", "), the translated sentence is obtained by using a back-translation procedure BID30 , i.e. by using the learned model to translate the source sentence to the target domain. In", "addition to these reconstruction objectives, we constrain the source and target sentence latent representations to have the same distribution using an adversarial regularization term, whereby the model tries to fool a discriminator which is simultaneously trained to identify the language of a given latent sentence representation BID8 . This", "procedure is then iteratively repeated, giving rise to translation models of increasing quality. To keep", "our approach fully unsupervised, we initialize our algorithm by using a naïve unsupervised translation model based on a word by word translation of sentences with a bilingual lexicon derived from the same monolingual data BID4 . As a result", ", and by only using monolingual data, we can encode sentences of both languages into the same feature space, and from there, we can also decode/translate in any of these languages; see FIG0 for an illustration.While not being able to compete with supervised approaches using lots of parallel resources, we show in section 4 that our model is able to achieve remarkable performance. For instance", ", on the WMT dataset we can achieve the same translation quality of a similar machine translation system trained with full supervision on 100,000 sentence pairs. On the Multi30K-Task1", "dataset we achieve a BLEU above 22 on all the language pairs, with up to 32.76 on English-French.Next, in section 2, we describe the model and the training algorithm. We then present experimental", "results in section 4. Finally, we further discuss", "related work in section 5 and summarize our findings in section 6.", "We presented a new approach to neural machine translation where a translation model is learned using monolingual datasets only, without any alignment between sentences or documents.", "The principle of our approach is to start from a simple unsupervised word-by-word translation model, and to iteratively improve this model based on a reconstruction loss, and using a discriminator to align latent distributions of both the source and the target languages.", "Our experiments demonstrate that our approach is able to learn effective translation models without any supervision of any sort." ]

[ 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.2916666567325592, 0.08695651590824127, 0.1599999964237213, 0.3265306055545807, 0.16326530277729034, 0.29032257199287415, 0.3333333134651184, 0.08510638028383255, 0.17777776718139648, 0.20895521342754364, 0.1538461446762085, 0.1818181723356247, 0.14814814925193787, 0.17910447716712952, 0.1304347813129425, 0.17543859779834747, 0.1395348757505417, 0, 0.09999999403953552, 0.11320754140615463, 0.09999999403953552, 0.18518517911434174, 0.1428571343421936, 0.0555555522441864, 0.09302324801683426, 0.1764705777168274, 0.11764705181121826, 0.19607841968536377, 0.11999999731779099, 0.1515151411294937, 0.0952380895614624, 0.24137930572032928, 0.24096384644508362, 0.26923075318336487, 0.20689654350280762, 0.05714285373687744, 0.054054051637649536, 0.31372547149658203, 0.2950819730758667, 0.2222222238779068 ]

[ "We propose a new unsupervised machine translation model that can learn without using parallel corpora; experimental results show impressive performance on multiple corpora and pairs of languages." ]

[ "We derive a new intrinsic social motivation for multi-agent reinforcement learning (MARL), in which agents are rewarded for having causal influence over another agent's actions, where causal influence is assessed using counterfactual reasoning.", "The reward does not depend on observing another agent's reward function, and is thus a more realistic approach to MARL than taken in previous work.", "We show that the causal influence reward is related to maximizing the mutual information between agents' actions.", "We test the approach in challenging social dilemma environments, where it consistently leads to enhanced cooperation between agents and higher collective reward.", "Moreover, we find that rewarding influence can lead agents to develop emergent communication protocols.", "Therefore, we also employ influence to train agents to use an explicit communication channel, and find that it leads to more effective communication and higher collective reward.", "Finally, we show that influence can be computed by equipping each agent with an internal model that predicts the actions of other agents.", "This allows the social influence reward to be computed without the use of a centralised controller, and as such represents a significantly more general and scalable inductive bias for MARL with independent agents.", "Deep reinforcement learning (RL) has made impressive progress on specific tasks with well-defined reward functions, but is still difficult to learn intelligent behavior that generalizes across multiple domains.", "Intrinsic motivation is a technique for solving this problem by developing general reward functions that encourage an agent to learn across a variety of tasks BID26 .", "Previous approaches to intrinsic motivation have broadly fallen into two categories: (1) curiosity, or a drive for novelty (e.g. BID17 BID24 ), and (2) empowerment, or a drive to be able to manipulate the environment (Klyubin et al., 2005) .We", "posit that this body of work has largely overlooked an important intrinsic motivation that is key to human learning: social interaction. Humans", "have remarkable social learning abilities; some authors suggest that it is social learning that has given rise to cultural evolution, and allowed us to achieve unprecedented progress and coordination on a massive scale BID31 BID11 . Others", "emphasize that our impressive capacity to learn from others far surpasses that of other animals, apes, and even other proto-human species BID10 BID9 Laland, 2017) . Therefore", ", we propose an intrinsic reward function designed for multi-agent RL (MARL), which awards agents for having a causal influence on other agents' actions. Causal influence", "is assessed using counterfactual reasoning; at each timestep, an agent simulates alternate, counterfactual actions that it could have taken, and assesses their effect on another agent's behavior. Actions that lead", "to relatively higher change in the other agent are considered to be highly influential and are rewarded. We show how this", "reward is related to maximizing the mutual information between agents' actions, and is thus a form of social empowerment. We hypothesize that", "rewarding influence may therefore encourage cooperation between agents. We also take inspiration", "from experiments in human cognition, showing that newborn infants are sensitive to correspondences between their own actions and the actions of other people, and use this to coordinate their behavior with others BID30 BID13 .To study our proposed social", "influence reward in the MARL setting, we adopt the Sequential Social Dilemmas (SSDs) of BID20 . These are challenging MA environments", "with a game-theoretic reward structure, similar to Prisoner's Dilemma. For each individual agent, 'defecting", "' (non-cooperative behavior) has the highest payoff. However, the collective reward will", "be better if all agents choose to cooperate. The paradoxical payoff structure of", "these tasks make achieving cooperative social dynamics• Finally, rather than computing social influence using a centralised training framework as in prior work (e.g. BID4 BID3 ), we extend the approach by attaching an internal Model of Other Agents (MOA) network to each agent and training it to predict the actions of every other agent. The agent can then simulate counterfactual", "actions and use its own internal MOA to predict how these will affect other agents, thus computing its own intrinsic influence reward.Using a MOA to predict and reward influence allows us to compute an intrinsic social reward by observing other agents' past actions, without a centralised controller, and without requiring access to another agent's reward function. We believe this is an important innovation", "over prior work (e.g. (Hughes et al., 2018; BID4 BID3 ). When we consider likely future applications", "of MARL, such as autonomous driving, it becomes apparent that centralised training or the sharing of reward functions are unrealistic assumptions, since autonomous vehicles are likely to be produced by a wide variety of organizations and institutions with mixed motivations. Rather, a social reward function which only", "depends on observing the behavior of agents acting in the environment, and which can give rise to coordinated, cooperative behavior, represents a more promising approach.", "The experiments above have demonstrated that an intrinsic social reward based on having causal influence on the actions of other agents consistently improves cooperation and leads to higher collective return in the MA social dilemmas under investigation.", "In some cases, the influence reward drove agents to learn an emergent communication protocol via their actions.", "This is compelling, and confirms the connection between maximizing influence and maximizing the mutual information between agents' actions.However, it is important to consider the limitations of the influence reward.", "Whether it will always give rise to cooperative behavior may depend on the specifics of the environment, task, and the trade-off between environmental and influence reward.", "Although influence is arguably necessary for cooperation (e.g. two agents cooperating to lift a box would have a high degree of influence between their actions), it may not be sufficient, in that it may be possible to influence another agent without helping it.", "For example, it is possible that agents could have gained influence in the tasks studied here by threatening to attack other agents with their fining beam.", "We believe this type of behavior did not emerge because communicating information represents the cheapest and most effective way to gain influence.", "Influencers do not have to sacrifice much in terms of their own environmental reward in order to communicate to other agents.Rewarding influence over an explicit communication channel may not be subject to this limitation, because influential communication may be inherently beneficial to the listener (at least in the case where listeners and speakers interact repeatedly).", "Since listeners can easily ignore communication messages if they do not help to obtain environmental reward, a speaker must transmit valuable information in order to gain influence through communication.", "There is no advantage to the speaker for communicating unreliably, because it would lose influence with the listener over time (although this is no longer guaranteed in one-shot interactions).", "Indeed, our results reveal that agents benefit from being influenced by (listening to) communication messages by obtaining higher individual reward, suggesting that the messages contain valuable information.", "Further, we found that the communication protocols learned via influence reward were more meaningful, and that the influence reward allowed agents to obtain higher collective return.", "Therefore, we suggest that influence could be a promising way to train emergent communication protocols in various settings.Finally, we have shown that influence can be computed by augmenting agents with an internal model that predicts the actions of other agents, and using this MOA model to simulate the effect of an agent's actions on others.", "This represents an important step forward in multi-agent intrinsic social motivation, because it implies that the influence reward can be computed without having access to another agent's reward function, or requiring a centralised controller." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.23728813230991364, 0.22641508281230927, 0.4000000059604645, 0.2745097875595093, 0.3255814015865326, 0.307692289352417, 0.31372547149658203, 0.33898305892944336, 0.14035087823867798, 0.25925925374031067, 0.1818181723356247, 0.1599999964237213, 0.19999998807907104, 0.18867923319339752, 0.37735849618911743, 0.14035087823867798, 0.2916666567325592, 0.3199999928474426, 0.19512194395065308, 0.25, 0.1666666567325592, 0.1395348757505417, 0.09999999403953552, 0.1904761791229248, 0.1927710771560669, 0.2666666507720947, 0, 0.19999998807907104, 0.3396226465702057, 0.4444444477558136, 0.3478260934352875, 0.2745097875595093, 0.307692289352417, 0.2461538463830948, 0.2222222238779068, 0.2745097875595093, 0.2702702581882477, 0.1428571343421936, 0.1818181723356247, 0.15094339847564697, 0.3921568691730499, 0.4324324131011963, 0.22580644488334656 ]

[ "We reward agents for having a causal influence on the actions of other agents, and show that this gives rise to better cooperation and more meaningful emergent communication protocols. " ]

[ "This work adopts the very successful distributional perspective on reinforcement learning and adapts it to the continuous control setting.", "We combine this within a distributed framework for off-policy learning in order to develop what we call the Distributed Distributional Deep Deterministic Policy Gradient algorithm, D4PG.", "We also combine this technique with a number of additional, simple improvements such as the use of N-step returns and prioritized experience replay.", "Experimentally we examine the contribution of each of these individual components, and show how they interact, as well as their combined contributions.", "Our results show that across a wide variety of simple control tasks, difficult manipulation tasks, and a set of hard obstacle-based locomotion tasks the D4PG algorithm achieves state of the art performance.", "The ability to solve complex control tasks with high-dimensional input and action spaces is a key milestone in developing real-world artificial intelligence.", "The use of reinforcement learning to solve these types of tasks has exploded following the work of the Deep Q Network (DQN) algorithm BID11 , capable of human-level performance on many Atari games.", "Similarly, ground breaking achievements have been made in classical games such as Go .", "However, these algorithms are restricted to problems with a finite number of discrete actions.In control tasks, commonly seen in the robotics domain, continuous action spaces are the norm.", "For algorithms such as DQN the policy is only implicitly defined in terms of its value function, with actions selected by maximizing this function.", "In the continuous control domain this would require either a costly optimization step or discretization of the action space.", "While discretization is perhaps the most straightforward solution, this can prove a particularly poor approximation in highdimensional settings or those that require finer grained control.", "Instead, a more principled approach is to parameterize the policy explicitly and directly optimize the long term value of following this policy.In this work we consider a number of modifications to the Deep Deterministic Policy Gradient (DDPG) algorithm BID9 .", "This algorithm has several properties that make it ideal for the enhancements we consider, which is at its core an off-policy actor-critic method.", "In particular, the policy gradient used to update the actor network depends only on a learned critic.", "This means that any improvements to the critic learning procedure will directly improve the quality of the actor updates.", "In this work we utilize a distributional BID0 version of the critic update which provides a better, more stable learning signal.", "Such distributions model the randomness due to intrinsic factors, among these is the inherent uncertainty imposed by function approximation in a continuous environment.", "We will see that using this distributional update directly results in better gradients and hence improves the performance of the learning algorithm.Due to the fact that DDPG is capable of learning off-policy it is also possible to modify the way in which experience is gathered.", "In this work we utilize this fact to run many actors in parallel, all feeding into a single replay table.", "This allows us to seamlessly distribute the task of gathering Authors contributed equally.", "The Deterministic Policy Gradient (DPG) algorithm BID19 upon which this work is based starts from a different set of ideas, namely the policy gradient theorem of BID22 .", "The deterministic policy gradient theorem builds upon this earlier approach, but replaces the stochastic policy with one that includes no randomness.", "This approach is particularly important because it had previously been believed that the deterministic policy gradient did not exist in a model-free setting.", "The form of this gradient is also interesting in that it does not require one to integrate over the action space, and hence may require less samples to learn.", "DPG was later built upon by BID9 who extended this algorithm and made use of a deep neural network as the function approximator, primarily as a mechanism for extending these results to work with vision-based inputs.", "Further, this entire endeavor lends itself very readily to an off-policy actorcritic architecture such that the actor's gradients depend only on derivatives through the learned critic.", "This means that by improving estimation of the critic one is directly able to improve the actor gradients.", "Most interestingly, there have also been recent attempts to distribute updates for the DDPG algorithm, (e.g. BID15 and more generally in this work we build on work of BID5 for implementing distributed actors.Recently, BID0 showed that the distribution over returns, whose expectation is the value function, obeys a distributional Bellman equation.", "Although the idea of estimating a distribution over returns has been revisited before BID21 BID13 , Bellemare et al. demonstrated that this estimation alone was enough to achieve state-of-the-art results on the Atari 2600 benchmarks.", "Crucially, this technique achieves these gains by directly improving updates for the critic.", "In this work we introduced the D4PG, or Distributed Distributional DDPG, algorithm.", "Our main contributions include the inclusion of a distributional updates to the DDPG algorithm, combined with the use of multiple distributed workers all writing into the same replay table.", "We also consider a number of other, smaller changes to the algorithm.", "All of these simple modifications contribute to the overall performance of the D4PG algorithm; the biggest performance gain of these simple changes is arguably the use of N -step returns.", "Interestingly we found that the use of priority was less crucial to the overall D4PG algorithm especially on harder problems.", "While the use of prioritization was definitely able to increase the performance of the D3PG algorithm, we found that it can also lead to unstable updates.", "This was most apparent in the manipulation tasks.Finally, as our results can attest, the D4PG algorithm is capable of state-of-the-art performance on a number of very difficult continuous control problems." ]

[ 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.17777776718139648, 0.4528301954269409, 0.20408162474632263, 0.12765957415103912, 0.3333333134651184, 0.08163265138864517, 0.178571417927742, 0, 0.25925925374031067, 0.07843136787414551, 0.2222222238779068, 0.1538461446762085, 0.29999998211860657, 0.19999998807907104, 0.1395348757505417, 0.13636362552642822, 0.21276594698429108, 0.12244897335767746, 0.1904761791229248, 0.08695651590824127, 0.09999999403953552, 0.2641509473323822, 0.08510638028383255, 0.11999999731779099, 0.1111111044883728, 0.09836065024137497, 0.1538461446762085, 0.13636362552642822, 0.18421052396297455, 0.16393442451953888, 0.09999999403953552, 0.1538461446762085, 0.1538461446762085, 0.25641024112701416, 0.1249999925494194, 0.260869562625885, 0.2448979616165161, 0.3214285671710968 ]

[ "We develop an agent that we call the Distributional Deterministic Deep Policy Gradient algorithm, which achieves state of the art performance on a number of challenging continuous control problems." ]

[ "State-action value functions (i.e., Q-values) are ubiquitous in reinforcement learning (RL), giving rise to popular algorithms such as SARSA and Q-learning.", "We propose a new notion of action value defined by a Gaussian smoothed version of the expected Q-value used in SARSA.", "We show that such smoothed Q-values still satisfy a Bellman equation, making them naturally learnable from experience sampled from an environment.", "Moreover, the gradients of expected reward with respect to the mean and covariance of a parameterized Gaussian policy can be recovered from the gradient and Hessian of the smoothed Q-value function.", "Based on these relationships we develop new algorithms for training a Gaussian policy directly from a learned Q-value approximator.", "The approach is also amenable to proximal optimization techniques by augmenting the objective with a penalty on KL-divergence from a previous policy.", "We find that the ability to learn both a mean and covariance during training allows this approach to achieve strong results on standard continuous control benchmarks.", "Model-free reinforcement learning algorithms often alternate between two concurrent but interacting processes: (1) policy evaluation, where an action value function (i.e., a Q-value) is updated to obtain a better estimate of the return associated with taking a specific action, and (2) policy improvement, where the policy is updated aiming to maximize the current value function.", "In the past, different notions of Q-value have led to distinct but important families of RL methods.", "For example, SARSA BID18 BID22 BID26 ) uses the expected Q-value, defined as the expected return of following the current policy.", "Q-learning BID28 ) exploits a hard-max notion of Q-value, defined as the expected return of following an optimal policy.", "Soft Q-learning BID7 and PCL BID14 both use a soft-max form of Q-value, defined as the future return of following an optimal entropy regularized policy.", "Clearly, the choice of Q-value function has a considerable effect on the resulting algorithm; for example, restricting the types of policies that can be expressed, and determining the type of exploration that can be naturally applied.In this work we introduce a new notion of action value: the smoothed action value functionQ π .", "Unlike previous notions, which associate a value with a specific action at each state, the smoothed Qvalue associates a value with a specific distribution over actions.", "In particular, the smoothed Q-value of a state-action pair (s, a) is defined as the expected return of first taking an action sampled from a normal distribution N (a, Σ(s)), centered at a, then following actions sampled from the current policy thereafter.", "In this way, the smoothed Q-value can also be interpreted as a Gaussian-smoothed or noisy version of the expected Q-value.We show that smoothed Q-values possess a number of interesting properties that make them attractive for use in RL algorithms.", "For one, the smoothed Q-values satisfy a single-step Bellman consistency, which allows bootstrapping to be used to train a function approximator.", "Secondly, for Gaussian policies, the standard optimization objective (expected return) can be expressed in terms of smoothed Q-values.", "Moreover, the gradient of this objective with respect to the mean and covariance of the Gaussian policy is equivalent to the gradient and the Hessian of the smoothed Q-value function, which allows one to derive updates to the policy parameters by having access to the derivatives of a sufficiently accurate smoothed Q-value function.This observation leads us to propose an algorithm called Smoothie, which in the spirit of (Deep) Deterministic Policy Gradient (DDPG) BID21 BID11 , trains a policy using the derivatives of a trained (smoothed) Q-value function, thus avoiding the high-variance of stochastic updates used in standard policy gradient algorithms BID29 BID10 .", "Unlike DDPG, which is well-known to have poor exploratory behavior BID7 , the approach we develop is able to utilize a non-deterministic Gaussian policy parameterized by both a mean and a covariance, thus allowing the policy to be exploratory by default and alleviating the need for excessive hyperparameter tuning.Furthermore, we show that Smoothie can be easily adapted to incorporate proximal policy optimization techniques by augmenting the objective with a penalty on KL-divergence from a previous version of the policy.", "The inclusion of a KL-penalty is not feasible in the standard DDPG algorithm, but we show that it is possible with our formulation, and it significantly improves stability and overall performance.", "On standard continuous control benchmarks, our results are competitive with or exceed state-of-the-art, especially for more difficult tasks in the low-data regime.", "We have presented a new Q-value function,Q π , that is a Gaussian-smoothed version of the standard expected Q-value, Q π .", "The advantage of usingQ π over Q π is that its gradient and Hessian possess an intimate relationship with the gradient of expected reward with respect to mean and covariance of a Gaussian policy.", "The resulting algorithm, Smoothie, is able to successfully learn both mean and covariance during training, leading to performance that can match or surpass that of DDPG, especially when incorporating a penalty on divergence from a previous policy.The success ofQ π is encouraging.", "Intuitively it may be argued that learningQ π is more sensible than learning Q π .", "The smoothed Q-values by definition make the true reward surface smoother, thus possibly easier to learn; moreover the smoothed Q-values have a more direct relationship with the expected discounted return objective.", "We encourage future work to further investigate these claims as well as techniques to apply the underlying motivations forQ π to other types of policies.A PROOF OF EQUATION FORMULA0 We note that similar identities for Gaussian integrals exist in the literature BID16 BID17 and point the reader to these works for further information.The specific identity we state may be derived using standard matrix calculus.", "We make use of the fact that DISPLAYFORM0 and for symmetric A, ∂ ∂A ||v|| DISPLAYFORM1 We omit s from Σ(s) in the following equations for succinctness.", "The LHS of FORMULA0 Meanwhile, towards tackling the RHS of FORMULA0 we note that DISPLAYFORM2 Thus we have DISPLAYFORM3" ]

[ 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.0555555522441864, 0.4375, 0.1818181723356247, 0.2631579041481018, 0.25806450843811035, 0.05882352590560913, 0.15789473056793213, 0.17241379618644714, 0.13793103396892548, 0.06451612710952759, 0.12903225421905518, 0.10810810327529907, 0.25925925374031067, 0.060606054961681366, 0.12244897335767746, 0.21276594698429108, 0.1249999925494194, 0.12903225421905518, 0.15189872682094574, 0.1111111044883728, 0.1463414579629898, 0, 0.375, 0.19512194395065308, 0.11764705181121826, 0.14814814925193787, 0.04999999701976776, 0.14492753148078918, 0.1621621549129486, 0.13793103396892548 ]

[ "We propose a new Q-value function that enables better learning of Gaussian policies." ]

[ "Interactive Fiction games are text-based simulations in which an agent interacts with the world purely through natural language.", "They are ideal environments for studying how to extend reinforcement learning agents to meet the challenges of natural language understanding, partial observability, and action generation in combinatorially-large text-based action spaces.", "We present KG-A2C, an agent that builds a dynamic knowledge graph while exploring and generates actions using a template-based action space.", "We contend that the dual uses of the knowledge graph to reason about game state and to constrain natural language generation are the keys to scalable exploration of combinatorially large natural language actions.", "Results across a wide variety of IF games show that KG-A2C outperforms current IF agents despite the exponential increase in action space size.", "Natural language communication has long been considered a defining characteristic of human intelligence.", "We are motivated by the question of how learning agents can understand and generate contextually relevant natural language in service of achieving a goal.", "In pursuit of this objective we study Interactive Fiction (IF) games, or text-adventures: simulations in which an agent interacts with the world purely through natural language-\"seeing\" and \"talking\" to the world using textual descriptions and commands.", "To progress in these games, an agent must generate natural language actions that are coherent, contextually relevant, and able to effect the desired change in the world.", "Complicating the problem of generating contextually relevant language in these games is the issue of partial observability: the fact that the agent never has access to the true underlying world state.", "IF games are structured as puzzles and often consist of an complex, interconnected web of distinct locations, objects, and characters.", "The agent needs to thus reason about the complexities of such a world solely through the textual descriptions that it receives, descriptions that are often incomplete.", "Further, an agent must be able to perform commonsense reasoning-IF games assume that human players possess prior commonsense and thematic knowledge (e.g. knowing that swords can kill trolls or that trolls live in dark places).", "Knowledge graphs provide us with an intuitive way of representing these partially observable worlds.", "Prior works have shown how using knowledge graphs aids in the twin issues of partial observability (Ammanabrolu & Riedl, 2019a) and commonsense reasoning (Ammanabrolu & Riedl, 2019b ), but do not use them in the context of generating natural language.", "To gain a sense for the challenges surrounding natural language generation, we need to first understand how large this space really is.", "In order to solve solve a popular IF game such as Zork1 it's necessary to generate actions consisting of up to five-words from a relatively modest vocabulary of 697 words recognized by Zork's parser.", "Even this modestly sized vocabulary leads to O(697 5 ) = 1.64 × 10 the structure required to further constrain our action space via our knowledge graph-and make the argument that the combination of these approaches allows us to generate meaningful natural language commands.", "Our contributions are as follows: We introduce an novel agent that utilizes both a knowledge graph based state space and template based action space and show how to train such an agent.", "We then conduct an empirical study evaluating our agent across a diverse set of IF games followed by an ablation analysis studying the effectiveness of various components of our algorithm as well as its overall generalizability.", "Remarkably we show that our agent achieves state-of-the-art performance on a large proportion of the games despite the exponential increase in action space size.", "Tabula rasa reinforcement learning offers an intuitive paradigm for exploring goal driven, contextually aware natural language generation.", "The sheer size of the natural language action space, however, has proven to be out of the reach of existing algorithms.", "In this paper we introduced KG-A2C, a novel learning agent that demonstrates the feasibility of scaling reinforcement learning towards natural language actions spaces with hundreds of millions of actions.", "The key insight to being able to efficiently explore such large spaces is the combination of a knowledge-graph-based state space and a template-based action space.", "The knowledge graph serves as a means for the agent to understand its surroundings, accumulate information about the game, and disambiguate similar textual observations while the templates lend a measure of structure that enables us to exploit that same knowledge graph for language generation.", "Together they constrain the vast space of possible actions into the compact space of sensible ones.", "An ablation study on Zork1 shows state-of-the-art performance with respect to any currently existing general reinforcement learning agent, including those with action spaces six orders of magnitude smaller than what we consider-indicating the overall efficacy of our combined state-action space.", "Further, a suite of experiments shows wide improvement over TDQN, the current state-of-the-art template based agent, across a diverse set of 26 human-made IF games covering multiple genres and game structures demonstrate that our agent is able to generalize effectively.", "A IMPLEMENTATION DETAILS" ]

[ 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.19607841968536377, 0.2950819730758667, 0.6792452931404114, 0.2711864411830902, 0.3272727131843567, 0.1304347813129425, 0.2857142686843872, 0.1515151411294937, 0.17241378128528595, 0.16949151456356049, 0.11764705181121826, 0.1428571343421936, 0.1538461446762085, 0.04255318641662598, 0.17910447716712952, 0.145454540848732, 0.06451612710952759, 0.19178082048892975, 0.29999998211860657, 0.1875, 0.2857142686843872, 0.19999998807907104, 0.15686273574829102, 0.3103448152542114, 0.2181818187236786, 0.260869562625885, 0.08695651590824127, 0.1690140813589096, 0.2535211145877838, 0 ]

[ "We present KG-A2C, a reinforcement learning agent that builds a dynamic knowledge graph while exploring and generates natural language using a template-based action space - outperforming all current agents on a wide set of text-based games." ]

[ "It is well-known that neural networks are universal approximators, but that deeper networks tend in practice to be more powerful than shallower ones.", "We shed light on this by proving that the total number of neurons m required to approximate natural classes of multivariate polynomials of n variables grows only linearly with n for deep neural networks, but grows exponentially when merely a single hidden layer is allowed.", "We also provide evidence that when the number of hidden layers is increased from 1 to k, the neuron requirement grows exponentially not with n but with n^{1/k}, suggesting that the minimum number of layers required for practical expressibility grows only logarithmically with n.", "Deep learning has lately been shown to be a very powerful tool for a wide range of problems, from image segmentation to machine translation.", "Despite its success, many of the techniques developed by practitioners of artificial neural networks (ANNs) are heuristics without theoretical guarantees.", "Perhaps most notably, the power of feedforward networks with many layers (deep networks) has not been fully explained.", "The goal of this paper is to shed more light on this question and to suggest heuristics for how deep is deep enough.It is well-known BID7 BID11 BID15 BID1 BID23 that neural networks with a single hidden layer can approximate any function under reasonable assumptions, but it is possible that the networks required will be extremely large.", "Recent authors have shown that some functions can be approximated by deeper networks much more efficiently (i.e. with fewer neurons) than by shallower ones.", "Often, these results admit one or more of the following limitations: \"existence proofs\" without explicit constructions of the functions in question; explicit constructions, but relatively complicated functions; or applicability only to types of network rarely used in practice.It is important and timely to extend this work to make it more concrete and actionable, by deriving resource requirements for approximating natural classes of functions using today's most common neural network architectures.", "BID17 recently proved that it is exponentially more efficient to use a deep network than a shallow network when Taylor-approximating the product of input variables.", "In the present paper, we move far beyond this result in the following ways:", "(i) we use standard uniform approximation instead of Taylor approximation,", "(ii) we show that the exponential advantage of depth extends to all general sparse multivariate polynomials, and", "(iii) we address the question of how the number of neurons scales with the number of layers.", "Our results apply to standard feedforward neural networks and are borne out by empirical tests.Our primary contributions are as follows:• It is possible to achieve arbitrarily close approximations of simple multivariate and univariate polynomials with neural networks having a bounded number of neurons (see §3).•", "Such polynomials are exponentially easier to approximate with deep networks than with shallow networks (see §4).•", "The power of networks improves rapidly with depth; for natural polynomials, the number of layers required is at most logarithmic in the number of input variables, where the base of the logarithm depends upon the layer width (see §5).", "We have shown how the power of deeper ANNs can be quantified even for simple polynomials.", "We have proved that arbitrarily good approximations of polynomials are possible even with a fixed number of neurons and that there is an exponential gap between the width of shallow and deep networks required for approximating a given sparse polynomial.", "For n variables, a shallow network requires size exponential in n, while a deep network requires at most linearly many neurons.", "Networks with a constant number k > 1 of hidden layers appear to interpolate between these extremes, following a curve exponential in n 1/k .", "This suggests a rough heuristic for the number of layers required for approximating simple functions with neural networks.", "For example, if we want no layers to have more than 2 10 neurons, say, then the minimum number of layers required grows only as log 10 n.", "To further improve efficiency using the O(n) constructions we have presented, it suffices to increase the number of layers by a factor of log 2 10 ≈ 3, to log 2 n.The key property we use in our constructions is compositionality, as detailed in BID24 .", "It is worth noting that as a consequence our networks enjoy the property of locality mentioned in , which is also a feature of convolutional neural nets.", "That is, each neuron in a layer is assumed to be connected only to a small subset of neurons from the previous layer, rather than the entirety (or some large fraction).", "In fact, we showed (e.g. Prop. 4.6) that there exist natural functions computable with linearly many neurons, with each neuron is connected to at most two neurons in the preceding layer, which nonetheless cannot be computed with fewer than exponentially many neurons in a single layer, no matter how may connections are used.", "Our construction can also be framed with reference to the other properties mentioned in : those of sharing (in which weights are shared between neural connections) and pooling (in which layers are gradually collapsed, as our construction essentially does with recursive combination of inputs).", "This paper has focused exclusively on the resources (neurons and synapses) required to compute a given function for fixed network depth.", "(Note also results of BID18 ; BID13 ; BID12 for networks of fixed width.)", "An important complementary challenge is to quantify the resources (e.g. training steps) required to learn the computation, i.e., to converge to appropriate weights using training data -possibly a fixed amount thereof, as suggested in Zhang et al. (2017) .", "There are simple functions that can be computed with polynomial resources but require exponential resources to learn (Shalev-Shwartz et al., 2017) .", "It is quite possible that architectures we have not considered increase the feasibility of learning.", "For example, residual networks (ResNets) BID14 and unitary nets (see e.g. BID0 BID16 ) are no more powerful in representational ability than conventional networks of the same size, but by being less susceptible to the \"vanishing/exploding gradient\" problem, it is far easier to optimize them in practice.", "We look forward to future work that will help us understand the power of neural networks to learn.", "Without loss of generality, suppose that r i > 0 for i = 1, . . . , n.", "Let X be the multiset in which x i occurs with multiplicity r i .We", "first show that n i=1 (r i + 1) neurons are sufficient to approximate p(x). Appendix", "A in Lin et al. (2017) demonstrates that for variables y 1 , . . . , y N , the product y 1 · · · · · y N can be Taylorapproximated as a linear combination of the 2 N functions σ(±y 1 ± · · · ± y d ).Consider", "setting y 1 , . . . , y d equal to the elements of multiset X. Then, we conclude that we can approximate p(x) as a linear combination of the functions σ(±y 1 ± · · · ± y d ). However,", "these functions are not all distinct: there are r i + 1 distinct ways to assign ± signs to r i copies of x i (ignoring permutations of the signs). Therefore", ", there are DISPLAYFORM0" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.3589743673801422, 0.23728813230991364, 0.11538460850715637, 0, 0.1621621549129486, 0.0555555522441864, 0.14705881476402283, 0.2380952388048172, 0.07894736528396606, 0.3414634168148041, 0, 0, 0.17142856121063232, 0, 0.17241379618644714, 0.42424240708351135, 0.08163265138864517, 0.11764705181121826, 0.3396226465702057, 0.1666666567325592, 0, 0.17142856121063232, 0.09090908616781235, 0, 0.1428571343421936, 0.043478257954120636, 0.1492537260055542, 0.0714285671710968, 0, 0.06451612710952759, 0, 0.10256409645080566, 0.060606054961681366, 0.12903225421905518, 0.22857142984867096, 0.05882352590560913, 0.0624999962747097, 0.11764705181121826, 0.039215680211782455, 0.0416666604578495, 0.04651162400841713, 0.09090908616781235 ]

[ "We prove that deep neural networks are exponentially more efficient than shallow ones at approximating sparse multivariate polynomials." ]

[ "Convolutional neural networks (CNNs) in recent years have made a dramatic impact in science, technology and industry, yet the theoretical mechanism of CNN architecture design remains surprisingly vague.", "The CNN neurons, including its distinctive element, convolutional filters, are known to be learnable features, yet their individual role in producing the output is rather unclear.", "The thesis of this work is that not all neurons are equally important and some of them contain more useful information to perform a given task.", "Hence, we propose to quantify and rank neuron importance, and directly incorporate neuron importance in the objective function under two formulations: (1) a game theoretical approach based on Shapley value which computes the marginal contribution of each filter; and (2) a probabilistic approach based on what-we-call, the importance switch using variational inference.", "Using these two methods we confirm the general theory that some of the neurons are inherently more important than the others.", "Various experiments illustrate that learned ranks can be readily useable for structured network compression and interpretability of learned features.", "Neural networks have achieved state-of-the art results in various cognition tasks, including image and speech recognition, machine translation, reinforcement learning (Fergus et al., 2003; Mnih et al., 2013; Gu et al., 2018) .", "Many of these applications involved CNNs which excel in particular in the vision tasks due to its ability to capture visual by means of convolution filters.", "Although the effectiveness of convolutional networks is unquestionable, the details of the architecture design and what particularly makes neural network work in detail remain highly uncertain.", "The experimental results roughly confirm that the accuracy of the network and representational capacity is correlated with the depth of the network He et al., 2016; Montufar et al., 2014) .", "Interestingly, the deeper architecture also become wider, although the link between width and network expressivity is questionable (Poole et al., 2016) and the choice of the number of neurons is rather discretionary.", "As a result the discussion about the network architecture often revolves around the numbers of filters and layers and their relative positioning, putting aside the conversation about the quality of the information that it contains.", "The increasing size of the network architectures have faced scrutiny that made claims that the networks are overparametrized raising two main concerns: heavy computational load and potential overfitting .", "In response to the need to build networks that are smaller yet accurate, a stream of research attempted to remove redundant units, compress the networks and design lighter architectures (Iandola et al., 2016; .", "A widespread approach to network reduction has been removing weights that are small or even close to zero (Han et al., 2015) .", "This line of research implicitly discerns that nodes with larger weights are more significant for learning task than the small weights.", "As a result, broadly speaking, this approach divides features between those that are useful which are kept and those which are insignificant and therefore discarded, forming a sort of binary approach.", "In this work, we would like to scrutinize the individual filters and form an explicit theory that states that the units in the network (both convolutional filters and nodes in fully connected layers) are not equally important when it comes to performing an inference task.", "The corollary of this thesis is that CNNs learn features in a discriminative way so that some of them carry more significance than others, and the knowledge about the input is not uniformly distributed among the CNN features.", "This theory is in line of research that adding more filters does not make the network more expressive since learning relevant information to the network has already been addressed by other filters.", "Given the proposed theory, we would like to make a step forward in gaining insight what the CNN learns and propose to extend the binary approach to form a quantifiable ranking of features.", "In other words, we attempt to estimate the importance of each feature compared to the others with particular focus on convolutional filters, which may be visualized.", "We introduce a theoretical framework to quantify how important each feature is through proposing a feature ranking method based on two different approaches.", "The first approach derives from the game theoretical concept of Shapley value (Shapley, 1953) , which assesses the importance of an individual in a group of neurons based on its marginal contribution to the group.", "The second method takes a probabilistic approach and introduces additional learnable parameters, which we call importance switches, that take real values and are trained by means of variational inference to give more weight to the important features.", "The extensive experimental results using these approaches indicate that some features are inherently more significant than others.", "The theoretical underpinnings of the feature rankings have further direct practical implications we explore.", "Firstly, the knowledge of the ranking allows to know which features directly impact the score of our method and consequently a more informed way of building an effective model.", "Thus, we are able to build a network around the the relevant features and discard the less relevant ones, effectively compressing the network achieving state-of-the-art results.", "Secondly and perhaps more significantly, the feature ranking of convolutional features provides more interpretable information about the network and places meaning on particular features in the context of a given task, thus casting light on the black box models.", "To achieve human interpretability, we visualize the most significant features which significantly show the significance of repeated and complementary features.", "In summary, this work suggests a theory that the learnable CNN features contain inherent hierarchy where some of the features are more significant than others.", "This multidisciplinary work which builds on top of probability and game theoretical concepts proposes two methods to produce feature ranking and select most important features in the CNN network.", "The striking observation is that the different methods lead to similar results and allow to distinguish important nodes with larger confidence.", "The ranking methods allow to build an informed way to build a slim network architecture where the significant nodes remain and unimportant nodes are discarded.", "A future search for further methods which allow to quantify the neuron importance is the next step to develop the understanding of the feature importance in CNNs." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0 ]

[ 0.13793103396892548, 0.21052631735801697, 0.1071428507566452, 0.21917808055877686, 0.1599999964237213, 0.08163265138864517, 0.06666666269302368, 0.18518517911434174, 0.18518517911434174, 0.145454540848732, 0.10344827175140381, 0.17241378128528595, 0.14035087823867798, 0.13114753365516663, 0.07547169178724289, 0.11764705181121826, 0.072727270424366, 0.20588234066963196, 0.1269841194152832, 0.20338982343673706, 0.2711864411830902, 0.145454540848732, 0.23076923191547394, 0.13114753365516663, 0.1818181723356247, 0, 0.04444444179534912, 0.2142857164144516, 0.19230768084526062, 0.16129031777381897, 0.20408162474632263, 0.07407406717538834, 0.4067796468734741, 0.31372547149658203, 0.2641509473323822, 0.22641508281230927 ]

[ "We propose CNN neuron ranking with two different methods and show their consistency in producing the result which allows to interpret what network deems important and compress the network by keeping the most relevant nodes." ]

[ "This work presents a modular and hierarchical approach to learn policies for exploring 3D environments.", "Our approach leverages the strengths of both classical and learning-based methods, by using analytical path planners with learned mappers, and global and local policies.", "Use of learning provides flexibility with respect to input modalities (in mapper), leverages structural regularities of the world (in global policies), and provides robustness to errors in state estimation (in local policies).", "Such use of learning within each module retains its benefits, while at the same time, hierarchical decomposition and modular training allow us to sidestep the high sample complexities associated with training end-to-end policies.", "Our experiments in visually and physically realistic simulated 3D environments demonstrate the effectiveness of our proposed approach over past learning and geometry-based approaches.", "Navigation is a critical task in building intelligent agents.", "Navigation tasks can be expressed in many forms, for example, point goal tasks involve navigating to a specific coordinates and semantic navigation involves finding path to a specific scene or object.", "Such tasks may need to be performed in known (already mapped) or unknown environments.", "Irrespective of the task or the setting, a core problem in navigation is exploration, i.e., how to efficiently visit as much of the environment.", "This is useful for pre-mapping in known environments, or actually solving tasks in known environments.", "Recent work from has used end-to-end learning to tackle this problem.", "Their motivation is three fold:", "a) learning provides flexibility to the choice of input modalities (classical systems rely on observing geometry through use of specialized sensors, while learning systems can infer geometry directly from RGB images),", "b) use of learning can improve robustness to errors in explicit state estimation, and", "c) learning can effectively leverage structural regularities of the real world, leading to more efficient behavior in previously unseen environments.", "This lead to their design of an end-to-end trained neural network based policy that processed raw sensory observations to directly output actions that the agent should execute.", "While use of learning for exploration is well motivated, casting the exploration problem as an end-to-end learning problem has its own drawbacks.", "Learning about mapping, state-estimation and path-planning purely from data in an end-to-end manner can be prohibitively expensive.", "Consequently, past end-to-end learning work for exploration from relies on use of imitation learning and many millions of frames of experience, but still performs worse than classical methods that don't require any training at all.", "This motivates our work.", "In this paper, we investigate alternate formulations of employing learning for exploration that retains the advantages that learning has to offer, but doesn't suffer from the drawbacks of full-blown end-to-end learning.", "Our key conceptual insight is that use of learning for leveraging structural regularities of indoor environments, robustness to state-estimation errors, and flexibility with respect to input modalities, happens at different time scales and can thus be factored out.", "This motivates use of learning in a modular and hierarchical fashion inside of what one may call a 'classical navigation pipeline'.", "This results in navigation policies that can work with raw sensory inputs such as RGB images, are robust to state estimation errors, and leverage regularities of real world layout.", "This results in extremely competitive performance over both geometry-based methods and recent learning-based methods; at the same time requiring a fraction of the number of samples.", "More specifically, our proposed exploration architecture comprises of a learned mapper (and pose estimator), a global policy, and a local policy, that are interfaced via the map and an analytical path planner.", "The learned mapper, together with the pose estimator, produces free space maps from input RGB images.", "The global policy consumes this free-space map and employs learning to exploit structural regularities in layout of real world environments to produce long-term goals.", "These long-term goals are used to generate short-term goals for the local policy (using a geometric path-planner).", "This local policy uses learning to directly map raw RGB images to actions that the agent should execute.", "Use of learning in mapper provides flexibility with respect to input modality, learned global policy can exploit regularities in layout of real world layout of environments, while learned local policies can use visual feedback to exhibit more robust behaviour.", "At the same time, hierarchical and modular design and use of analytical planning, significantly cuts down the search space during training, leading to better performance as well as sample efficient learning.", "We demonstrate our proposed approach in visually and physically realistic simulators for the task of geometric exploration (visit as much area as possible).", "We work with the Habitat simulator from Savva et al. (2019) .", "While Habitat is already visually realistic (it uses real world scans from Chang et al. (2017) ; Xia et al. (2018) as environments), we improve its physical realism by using actuation and odometry sensor noise models, that we collected by conducting physical experiments on a real mobile robot.", "Our experiments and ablations in this realistic simulation reveal the effectiveness of our proposed approach for the task of exploration.", "A straight-forward modification of our method also tackles point-goal navigation tasks, and won the AI Habitat challenge at CVPR2019 across all tracks.", "In this paper, we proposed a modular navigational model which leverages the strengths of classical and learning-based navigational methods.", "We show that the proposed model outperforms prior methods on both Exploration and PointGoal tasks and shows strong generalization across domains, goals, and tasks.", "In future, the proposed model can be extended to complex semantic tasks such as Semantic Goal Navigation and Embodied Question Answering by using a semantic Mapper which creates multi-channel map capturing semantic properties of the objects in the environment.", "The model can also be combined with prior work on Localization to relocalize in a previously created map for efficient navigation in subsequent episodes." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.8148148059844971, 0.1764705777168274, 0.10256409645080566, 0.23255813121795654, 0.23529411852359772, 0, 0.1538461446762085, 0.1538461446762085, 0.05714285373687744, 0.1599999964237213, 0.08695651590824127, 0, 0.05128204822540283, 0.1538461446762085, 0.1249999925494194, 0.05405404791235924, 0.06451612710952759, 0.06896550953388214, 0.09090908616781235, 0, 0.10526315122842789, 0.12765957415103912, 0.19354838132858276, 0.1463414579629898, 0.0555555522441864, 0.04999999701976776, 0, 0.17142856121063232, 0.1428571343421936, 0.06896550953388214, 0.09090908616781235, 0.20000000298023224, 0.1764705777168274, 0, 0.03703703358769417, 0.19999998807907104, 0.11764705181121826, 0.13333332538604736, 0.060606054961681366, 0.08510638028383255, 0.11428570747375488 ]

[ "A modular and hierarchical approach to learn policies for exploring 3D environments." ]

[ "Deep Learning for Computer Vision depends mainly on the source of supervision.", "Photo-realistic simulators can generate large-scale automatically labeled synthetic data, but introduce a domain gap negatively impacting performance.", "We propose a new unsupervised domain adaptation algorithm, called SPIGAN, relying on Simulator Privileged Information (PI) and Generative Adversarial Networks (GAN).", "We use internal data from the simulator as PI during the training of a target task network.", "We experimentally evaluate our approach on semantic segmentation.", "We train the networks on real-world Cityscapes and Vistas datasets, using only unlabeled real-world images and synthetic labeled data with z-buffer (depth) PI from the SYNTHIA dataset.", "Our method improves over no adaptation and state-of-the-art unsupervised domain adaptation techniques.", "Learning from as little human supervision as possible is a major challenge in Machine Learning.", "In Computer Vision, labeling images and videos is the main bottleneck towards achieving large scale learning and generalization.", "Recently, training in simulation has shown continuous improvements in several tasks, such as optical flow BID32 , object detection BID31 BID52 BID47 BID36 , tracking BID10 , pose and viewpoint estimation BID44 BID34 BID46 , action recognition BID9 , and semantic segmentation BID15 BID39 BID38 .", "However, large domain gaps between synthetic and real domains remain as the main handicap of this type of strategies.", "This is often addressed by manually labeling some amount of real-world target data to train the model on mixed synthetic and real-world labeled data (supervised domain adaptation).", "In contrast, several recent unsupervised domain adaptation algorithms have leveraged the potential of Generative Adversarial Networks (GANs) BID14 for pixel-level adaptation in this context BID1 BID45 .", "These methods often use simulators as black-box generators of (x,", "y) input / output training samples for the desired task.Our main observation is that simulators internally know a lot more about the world and how the scene is formed, which we call Privileged Information (PI).", "This Privileged Information includes physical properties that might be useful for learning.", "This additional information z is not available in the real-world and is, therefore, generally ignored during learning.", "In this paper, we propose a novel adversarial learning algorithm, called SPIGAN, to leverage Simulator PI for GAN-based unsupervised learning of a target task network from unpaired unlabeled real-world data.We jointly learn four different networks:", "(i) a generator G (to adapt the pixel-level distribution of synthetic images to be more like real ones),", "(ii) a discriminator D (to distinguish adapted and real images),", "(iii) a task network T (to predict the desired label y from image x), and (iv) a privileged network P trained on both synthetic images x and adapted ones G(x) to predict their associated privileged information z.", "Our main contribution is a new method to leverage PI from a simulator via the privileged network P , which acts as an auxiliary task and regularizer to the task network T , the main output of our SPIGAN learning algorithm.We evaluate our approach on semantic segmentation in urban scenes, a challenging real-world task.", "We use the standard Cityscapes BID6 and Vistas BID33 datasets as target real-world data (without using any of the training labels) and SYNTHIA BID39 as simulator output.", "Although our method applies to any kind of PI that can be predicted via a deep network (optical flow, instance segmentation, object detection, material properties, forces, ...), we consider one of the most common and simple forms of PI available in any simulator: depth from its z-buffer.", "We show that SPIGAN can successfully learn a semantic segmentation network T using no real-world labels, partially bridging the sim-to-real gap (see Figure 1 ).", "SPIGAN also outperforms related state-of-the-art unsupervised domain adaptation methods.The rest of the paper is organized as follows.", "Section 2 presents a brief review of related works.", "Section 3 presents our SPIGAN unsupervised domain adaptation algorithm using simulator privileged information.", "We report our quantitative experiments on semantic segmentation in Section 4, and conclude in Section 5.", "In this section we present our evaluation of the SPIGAN algorithm in the context of adapting a semantic segmentation network from SYNTHIA to Cityscapes.", "Depth maps from SYNTHIA are used as PI in the proposed algorithm.We compare our results to several state-of-art domain adaptation algorithms, including FCNs in the wild (FCNs wild) BID21 , Curriculum DA (CDA) , Learning from synthetic data (LSD) BID42 , and Class-balanced Self-Training (CBST) BID59 .Quantitative", "results for these methods are shown in Table 1 for the semantic segmentation task on the target domain of Cityscapes (validation set). As reference", "baselines, we include results training only on source images and non-adapted labels. We also provide", "our algorithm performance without the PI for comparison (i.e., γ = 0 in Eq. 1, named \"SPIGAN-no-PI\").Results show that", "on Cityscapes SPIGAN achieves state-of-the-art semantic segmentation adaptation in terms of mean IoU. A finer analysis", "of the results attending to individual classes suggests that the use of PI helps to estimate layout-related classes such as road and sidewalk and object-related classes such as person, rider, car, bus and motorcycle. SPIGAN achieves", "an improvement of 3% in 320 × 640, 1.0% in 512 × 1024, in mean IoU with respect to the non-PI method. This improvement", "is thanks to the regularization provided by P (x; θ P ) during training, which decreases the number of artifacts as shown in Figure 5 . This comparison,", "therefore, confirms our main contribution: a general approach to leveraging synthetic data and PI from the simulator to improve generalization performance across the sim-to-real domain gap.", "We present SPIGAN, a novel method for leveraging synthetic data and Privileged Information (PI) available in simulated environments to perform unsupervised domain adaptation of deep networks.", "Our approach jointly learns a generative pixel-level adaptation network together with a target task network and privileged information models.", "We showed that our approach is able to address large domain gaps between synthetic data and target real-world domains, including for challenging realworld tasks like semantic segmentation of urban scenes.", "For future work, we plan to investigate SPIGAN applied to additional tasks, with different types of PI that can be obtained from simulation." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.06451612710952759, 0.1111111044883728, 0.19999998807907104, 0.17142856121063232, 0.14814814925193787, 0.1395348757505417, 0.2666666507720947, 0.1249999925494194, 0, 0.06896550953388214, 0.05405404791235924, 0.045454539358615875, 0.1818181723356247, 0, 0.07692307233810425, 0.06451612710952759, 0.0555555522441864, 0.18867924809455872, 0.05405404791235924, 0.06896550953388214, 0.19607841968536377, 0.2222222238779068, 0.09302324801683426, 0.09677419066429138, 0.22727271914482117, 0.1621621549129486, 0.0714285671710968, 0.4375, 0.12121211737394333, 0.19512194395065308, 0.09677419066429138, 0.19512194395065308, 0, 0.04878048226237297, 0.17142856121063232, 0, 0.09999999403953552, 0, 0.23255813121795654, 0.2666666507720947, 0.277777761220932, 0.16326530277729034, 0.09756097197532654 ]

[ "An unsupervised sim-to-real domain adaptation method for semantic segmentation using privileged information from a simulator with GAN-based image translation." ]

[ "Adversarial training is one of the main defenses against adversarial attacks.", "In this paper, we provide the first rigorous study on diagnosing elements of large-scale adversarial training on ImageNet, which reveals two intriguing properties. \n\n", "First, we study the role of normalization.", "Batch normalization (BN) is a crucial element for achieving state-of-the-art performance on many vision tasks, but we show it may prevent networks from obtaining strong robustness in adversarial training.", "One unexpected observation is that, for models trained with BN, simply removing clean images from training data largely boosts adversarial robustness, i.e., 18.3%.", "We relate this phenomenon to the hypothesis that clean images and adversarial images are drawn from two different domains.", "This two-domain hypothesis may explain the issue of BN when training with a mixture of clean and adversarial images, as estimating normalization statistics of this mixture distribution is challenging.", "Guided by this two-domain hypothesis, we show disentangling the mixture distribution for normalization, i.e., applying separate BNs to clean and adversarial images for statistics estimation, achieves much stronger robustness.", "Additionally, we find that enforcing BNs to behave consistently at training and testing can further enhance robustness.\n\n", "Second, we study the role of network capacity.", "We find our so-called \"deep\" networks are still shallow for the task of adversarial learning.", "Unlike traditional classification tasks where accuracy is only marginally improved by adding more layers to \"deep\" networks (e.g., ResNet-152), adversarial training exhibits a much stronger demand on deeper networks to achieve higher adversarial robustness.", "This robustness improvement can be observed substantially and consistently even by pushing the network capacity to an unprecedented scale, i.e., ResNet-638. \n", "Adversarial attacks (Szegedy et al., 2014) can mislead neural networks to make wrong predictions by adding human imperceptible perturbations to input data.", "Adversarial training (Goodfellow et al., 2015) is shown to be an effective method to defend against such attacks, which trains neural networks on adversarial images that are generated on-the-fly during training.", "Later works further improve robustness of adversarially trained models by mitigating gradient masking (Tramèr et al., 2018) , imposing logits pairing (Kannan et al., 2018) , denoising at feature space (Xie et al., 2019b) , etc.", "However, these works mainly focus on justifying the effectiveness of proposed strategies and apply inconsistent pipelines for adversarial training, which leaves revealing important elements for training robust models still a missing piece in current adversarial research.", "In this paper, we provide the first rigorous diagnosis of different adversarial learning strategies, under a unified training and testing framework, on the large-scale ImageNet dataset (Russakovsky et al., 2015) .", "We discover two intriguing properties of adversarial training, which are essential for training models with stronger robustness.", "First, though Batch Normalization (BN) (Ioffe & Szegedy, 2015) is known as a crucial component for achieving state-of-the-arts on many vision tasks, it may become a major obstacle for securing robustness against strong attacks in the context of adversarial training.", "By training such networks adversarially with different strategies, e.g., imposing logits pairing (Kannan et al., 2018) , we observe an unexpected phenomenon -removing clean images from training data is the most effective way for boosting model robustness.", "We relate this phenomenon to the conjecture that clean images and adversarial images are drawn from two different domains.", "This two-domain hypothesis may explain the limitation of BN when training with a mixture of clean and adversarial images, as estimating normalization statistics on this mixture distribution is challenging.", "We further show that adversarial training without removing clean images can also obtain strong robustness, if the mixture distribution is well disentangled at BN by constructing different mini-batches for clean images and adversarial images to estimate normalization statistics, i.e., one set of BNs exclusively for adversarial images and another set of BNs exclusively for clean images.", "An alternative solution to avoiding mixture distribution for normalization is to simply replace all BNs with batch-unrelated normalization layers, e.g., group normalization (Wu & He, 2018) , where normalization statistics are estimated on each image independently.", "These facts indicate that model robustness is highly related to normalization in adversarial training.", "Furthermore, additional performance gain is observed via enforcing consistent behavior of BN during training and testing.", "Second, we find that our so-called \"deep\" networks (e.g., are still shallow for the task of adversarial learning, and simply going deeper can effectively boost model robustness.", "Experiments show that directly adding more layers to \"deep\" networks only marginally improves accuracy for traditional image classification tasks.", "In contrast, substantial and consistent robustness improvement is witnessed even by pushing the network capacity to an unprecedented scale, i.e., ResNet-638.", "This phenomenon suggests that larger networks are encouraged for the task of adversarial learning, as the learning target, i.e., adversarial images, is a more complex distribution than clean images to fit.", "In summary, our paper reveals two intriguing properties of adversarial training: (1) properly handling normalization is essential for obtaining models with strong robustness; and (2) our so-called \"deep\" networks are still shallow for the task of adversarial learning.", "We hope these findings can benefit future research on understanding adversarial training and improving adversarial robustness.", "In this paper, we reveal two intriguing properties of adversarial training at scale: (1) conducting normalization in the right manner is essential for training robust models on large-scale datasets like ImageNet; and (2) our so-called \"deep\" networks are still shallow for the task of adversarial learning.", "Our discoveries may also be inherently related to our two-domain hypothesis -clean images and adversarial images are drawn from different distributions.", "We hope these findings can facilitate fellow researchers for better understanding of adversarial training as well as further improvement of adversarial robustness." ]

[ 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.2857142686843872, 0.3636363446712494, 0.11764705181121826, 0.1538461446762085, 0.1111111044883728, 0.0714285671710968, 0.1666666567325592, 0.04999999701976776, 0.0714285671710968, 0.1111111044883728, 0.1599999964237213, 0.1395348757505417, 0, 0, 0.14999999105930328, 0.05128204822540283, 0.1818181723356247, 0.3499999940395355, 0.2222222238779068, 0.1666666567325592, 0.04255318641662598, 0.0714285671710968, 0.21621620655059814, 0.11320754140615463, 0.04651162400841713, 0.1666666567325592, 0.1538461446762085, 0.10256409645080566, 0, 0, 0.09756097197532654, 0.09090908616781235, 0.23999999463558197, 0.19607843458652496, 0.06666666269302368, 0.20689654350280762 ]

[ "The first rigor diagnose of large-scale adversarial training on ImageNet" ]

[ "The gradient of a deep neural network (DNN) w.r.t. the input provides information that can be used to explain the output prediction in terms of the input features and has been widely studied to assist in interpreting DNNs.", " In a linear model (i.e., $g(x)=wx+b$), the gradient corresponds solely to the weights $w$.", "Such a model can reasonably locally linearly approximate a smooth nonlinear DNN, and hence the weights of this local model are the gradient.", "The other part, however, of a local linear model, i.e., the bias $b$, is usually overlooked in attribution methods since it is not part of the gradient.", "In this paper, we observe that since the bias in a DNN also has a non-negligible contribution to the correctness of predictions, it can also play a significant role in understanding DNN behaviors.", "In particular, we study how to attribute a DNN's bias to its input features.", "We propose a backpropagation-type algorithm ``bias back-propagation (BBp)'' that starts at the output layer and iteratively attributes the bias of each layer to its input nodes as well as combining the resulting bias term of the previous layer.", "This process stops at the input layer, where summing up the attributions over all the input features exactly recovers $b$.", "Together with the backpropagation of the gradient generating $w$, we can fully recover the locally linear model $g(x)=wx+b$.", "Hence, the attribution of the DNN outputs to its inputs is decomposed into two parts, the gradient $w$ and the bias attribution, providing separate and complementary explanations.", "We study several possible attribution methods applied to the bias of each layer in BBp.", "In experiments, we show that BBp can generate complementary and highly interpretable explanations of DNNs in addition to gradient-based attributions.", "Deep neural networks (DNNs) have produced good results for many challenging problems in computer vision, natural language processing, and speech processing.", "Deep learning models, however, are usually designed using fairly high-level architectural decisions, leading to a final model that is often seen as a difficult to interpret black box.", "DNNs are a highly expressive trainable class of non-linear functions, utilizing multi-layer architectures and a rich set of possible hidden non-linearities, making interpretation by a human difficult.", "This restricts the reliability and usability of DNNs especially in mission-critical applications where a good understanding of the model's behavior is necessary.The gradient is a useful starting point for understanding and generating explanations for the behavior of a complex DNN.", "Having the same dimension as the input data, the gradient can reflect the contribution to the DNN output of each input dimension.", "Not only does the gradient yield attribution information for every data point, but also it helps us understand other aspects of DNNs, such as the highly celebrated adversarial examples and defense methods against such attacks BID13 .When", "a model is linear, the gradient recovers the weight vector. Since", "a linear model locally approximates any sufficiently smooth non-linear model, the gradient can also be seen as the weight vector of that local linear model for a given DNN at a given data point. For a", "piecewise linear DNN (e.g., a DNN with activation functions such as ReLU, LeakyReLU, PReLU, and hard tanh) the gradient is exactly the weights of the local linear model 1 .Although", "the gradient of a DNN has been shown to be helpful in understanding the behavior of a DNN, the other part of the locally linear model, i.e., the bias term, to the best of our knowledge, has not been studied explicitly and is often overlooked. If only", "considering one linear model within a small region, the bias, as a scalar, seems to contain less information than the weight vector. However", ", this scalar is the result of complicated processing of bias terms over every neuron and every layer based on the activations, the non-linearity functions, as well as the weight matrices of the network. Uncovering", "the bias's nature could potentially reveal a rich vein of attribution information complementary to the gradient. For classification", "tasks, it can be the case that the gradient part of the linear model contributes to only a negligible portion of the target label's output probability (or even a negative logit value), and only with a large bias term does the target label's probability becomes larger than that of other labels to result in the correct prediction (see Sec 5). In our empirical experiments", "TAB0 , using only the bias term of the local linear models achieves 30-40% of the performance of the complete DNN, thus indicating that the bias term indeed plays a substantial role in the mechanisms of a DNN.In this paper, we unveil the information embedded in the bias term by developing a general bias attribution framework that distributes the bias scalar to every dimension of the input data. We propose a backpropagation-type", "algorithm called \"bias backpropagation (BBp)\" to send and compute the bias attribution from the output and higher-layer nodes to lower-layer nodes and eventually to the input features, in a layer-by-layer manner. Specifically, BBp utilizes a recursive", "rule to assign the bias attribution on each node of layer to all the nodes on layer − 1, while the bias attribution on each node of layer − 1 is composed of the attribution sent from the layer below and the bias term incurred in layer − 1. The sum of the attributions over all input", "dimensions produced by BBp exactly recovers the bias term in the local linear model representation of the DNN at the given input point. In experiments, we visualize the bias attribution", "results as images on a DNN trained for image classification. We show that bias attribution can highlight essential", "features that are complementary from what the gradient-alone attribution methods favor." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.4000000059604645, 0.1463414579629898, 0.17391303181648254, 0.19230768084526062, 0.22641508281230927, 0.25641024112701416, 0.2857142686843872, 0.1860465109348297, 0.0952380895614624, 0.2857142686843872, 0.24390242993831635, 0.52173912525177, 0.1702127605676651, 0.07692307233810425, 0.23999999463558197, 0.25, 0.1904761791229248, 0.13114753365516663, 0.1111111044883728, 0.1090909019112587, 0.14814814925193787, 0.2222222238779068, 0.12765957415103912, 0.18867923319339752, 0.23255813121795654, 0.18666666746139526, 0.23999999463558197, 0.25925925374031067, 0.28070175647735596, 0.23529411852359772, 0.09090908616781235, 0.1621621549129486 ]

[ "Attribute the bias terms of deep neural networks to input features by a backpropagation-type algorithm; Generate complementary and highly interpretable explanations of DNNs in addition to gradient-based attributions." ]

[ "This paper presents a method to autonomously find periodicities in a signal.", "It is based on the same idea of using Fourier Transform and autocorrelation function presented in Vlachos et al. 2005.", "While showing interesting results this method does not perform well on noisy signals or signals with multiple periodicities.", "Thus, our method adds several new extra steps (hints clustering, filtering and detrending) to fix these issues.", "Experimental results show that the proposed method outperforms the state of the art algorithms.", "A time series is defined by its 3 main components : the trend component, the periodic component and the random component.", "Trend analysis and prediction are topics that have been greatly studied BID10 and will not be treated in the article, therefore every time series will be assumed stationary regarding its mean and variance, so this study focus the periodic component.", "The ability to detect and find the main characteristic of this component is not as easy as the trend component.", "Yet, the ability to detect periodicities in a time series is essential to make precise forecasts.A periodicity is a pattern in a time series that occurs at regular time intervals.", "More precisely, the time series is said cyclical, if the time intervals at which the pattern repeats itself can't be precisely defined and is not constant.", "On the opposite, there are seasonal time series in which the pattern repeats itself at constant and well defined time intervals.", "Thus, cyclical patterns are more difficult to detect due to their inconsistency and the fact that they usually repeat themselves over large periods of time and therefore require more data to be identified.", "Nevertheless, seasonal patterns are very common in time series such as those related to human behaviour which usually have periodicities like hours and calendar (time of day, day of week, month of year).", "This kind of feature is well known and can be easily tested to see if they are beneficial or not.", "Unfortunately, when it comes to time series related to other phenomenons, the periodicities are not trivially found.", "For instance, tides level are multi-periodic time series correlated to both moon cycles and sun cycles; and females menstrual cycles are related to hormonal changes.", "The ability to detect periodicity in time series is fundamental when it comes to forecasting BID5 .", "Once a periodic pattern has been detected, numerous techniques can be used to model this later and improve forecasts BID1 .", "However, periodicities detection is not easy and has been greatly studied in the existing literature, but most of current techniques are unable to detect periodicities without the need of preprocessing data BID12 or have trouble detecting multiple periodicities BID11 .", "This paper is organised as follow: we first present the Fourier transform and the Autoperiod algorithm BID11 used to detect periodicities in a signal.", "Then we propose a new fully automated method, named Clustered Filtered Detrended Autoperiod (CFD-Autoperiod), which also combines the advantages of frequency domain and time domain while being robust to noise and able to handle multi periodicities.", "Noise robustness is achieved using a density clustering on hints provided by the frequency analysis.", "Multi-periodicities are more precisely detected by both using detrending and filtering.", "Finally, we demonstrate that CFD-Autoperiod outperforms previous methods." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.625, 0.1463414579629898, 0.15789473056793213, 0.21052631735801697, 0.060606054961681366, 0.051282044500112534, 0.0714285671710968, 0.15789473056793213, 0.1818181723356247, 0.04651162400841713, 0.09999999403953552, 0.07999999821186066, 0.1538461446762085, 0.1463414579629898, 0.10810810327529907, 0.0952380895614624, 0.1111111044883728, 0.1463414579629898, 0.178571417927742, 0.3181818127632141, 0.14814814925193787, 0.1111111044883728, 0.1249999925494194, 0 ]

[ "This paper presents a method to autonomously find multiple periodicities in a signal, using FFT and ACF and add three news steps (clustering/filtering/detrending)" ]

[ "We present an adversarial exploration strategy, a simple yet effective imitation learning scheme that incentivizes exploration of an environment without any extrinsic reward or human demonstration.", "Our framework consists of a deep reinforcement learning (DRL) agent and an inverse dynamics model contesting with each other.", "The former collects training samples for the latter, and its objective is to maximize the error of the latter.", "The latter is trained with samples collected by the former, and generates rewards for the former when it fails to predict the actual action taken by the former.", "In such a competitive setting, the DRL agent learns to generate samples that the inverse dynamics model fails to predict correctly, and the inverse dynamics model learns to adapt to the challenging samples.", "We further propose a reward structure that ensures the DRL agent collects only moderately hard samples and not overly hard ones that prevent the inverse model from imitating effectively.", "We evaluate the effectiveness of our method on several OpenAI gym robotic arm and hand manipulation tasks against a number of baseline models.", "Experimental results show that our method is comparable to that directly trained with expert demonstrations, and superior to the other baselines even without any human priors.", "Over the past decade, imitation learning (IL) has been successfully applied to a wide range of domains, including robot learning BID7 BID20 , autonomous navigation BID4 BID19 , manipulation tasks BID11 BID17 , and self-driving cars BID5 .", "Traditionally, IL aims to train an imitator to learn a control policy π only from expert demonstrations.", "The imitator is typically presented with multiple demonstrations during the training phase, with an aim to distill them into π.", "To learn π effectively and efficiently, a large set of high-quality demonstrations are necessary.", "This is especially prevalent in current state-of-the-art IL algorithms, such as dataset aggregation (DAgger) BID18 and generative adversarial imitation learning (GAIL) BID9 .", "Although these approaches have been the dominant algorithms in IL, a major bottleneck for them is their reliance on high-quality demonstrations, which often require extensive supervision from human experts.", "In addition, a serious flaw in the learned policy π is its tendency to overfit to demonstration data, preventing it from generalizing to new ones.", "To overcome the aforementioned challenges in IL, a number of methods have been investigated to enhance the generalizability and data efficiency, or reduce the degree of human supervision.", "Initial efforts in this direction were based on the idea of meta learning BID6 BID8 Yu et al., 2018) , in which the imitator is trained from a meta learner that is able to quickly learn a new task with only a few set of demonstrations.", "However, such schemes still require training the meta-learner with tremendous amount of time and demonstration data, leaving much room for improvement.", "Thus, a rapidly-growing body of literature based on the concept of using forward/inverse dynamics models to learn π within an environment in a self-supervised fashion BID0 BID11 BID13 has emerged in the past few years.", "One key advantage of the concept is that it provides an autonomous way for preparing training data, removing the need of human intervention.", "In this paper, we call it self-supervised IL.Self-supervised IL allows an imitator to collect training data by itself instead of using predefined extrinsic reward functions or expert supervision during training.", "It only needs demonstration during inference, drastically decreasing the time and effort required from human experts.", "Although the core principles of self-supervised IL are straightforward and have been exploited in many fields BID0 BID11 BID12 , recent research efforts have been dedicated to addressing the challenges of multi-modality and multi-step planning.", "For example, the use of forward consistency loss and forward regularizer have been extensively investigated to enhance the task performance of the imitator BID0 BID13 .", "This becomes especially essential when the lengths of trajectories grow and demonstration samples are sparse, as multiple paths may co-exist to lead the imitator from its initial observation to the goal observation.", "The issue of multi-step planning has also drawn a lot of attention from researchers, and is usually tackled by recurrent neural networks (RNNs) and step-by-step demonstrations BID11 BID13 .", "The above self-supervised IL approaches report promising results, however, most of them are limited in applicability due to several drawbacks.", "First, traditional methods of data collection are usually inefficient and time-consuming.", "Inefficient data collection results in poor exploration, giving rise to a degradation in robustness to varying environmental conditions (e.g., noise in motor control) and generalizability to difficult tasks.", "Second, human bias in data sampling range tailored to specific interesting configurations is often employed BID0 BID11 .", "Although a more general exploration strategy called curiosity-driven exploration was later proposed in BID12 , it focuses only on exploration in states novel to the forward dynamics model, rather than those directly influential to the inverse dynamics model.", "Furthermore, it does not discuss the applicability to continuous control domains, and fails in high dimensional action spaces according to our experiments in Section 4.", "Unlike the approaches discussed above, we do not propose to deal with multi-modality or multi-step planning.", "Instead, we focus our attention on improving the overall quality of the collected samples in the context of self-supervised IL.", "This motivates us to equip the model with the necessary knowledge to explore the environment in an efficient and effective fashion.In this paper, we propose a straightforward and efficient self-supervised IL scheme, called adversarial exploration strategy, which motivates exploration of an environment in a self-supervised manner (i.e., without any extrinsic reward or human demonstration).", "Inspired by Pinto et al. (2017) ; BID23 ; BID24 , we implement the proposed strategy by jointly training a deep reinforcement learning (DRL) agent and an inverse dynamics model competing with each other.", "The former explores the environment to collect training data for the latter, and receives rewards from the latter if the data samples are considered difficult.", "The latter is trained with the training data collected by the former, and only generates rewards when it fails to predict the true actions performed by the former.", "In such an adversarial setting, the DRL agent is rewarded only for the failure of the inverse dynamics model.", "Therefore, the DRL agent learns to sample hard examples to maximize the chances to fail the inverse dynamics model.", "On the other hand, the inverse dynamics model learns to be robust to the hard examples collected by the DRL agent by minimizing the probability of failures.", "As a result, as the inverse dynamics model becomes stronger, the DRL agent is also incentivized to search for harder examples to obtain rewards.", "Overly hard examples, however, may lead to biased exploration and cause instability of the learning process.", "In order to stabilize the learning curve of the inverse dynamics model, we further propose a reward structure such that the DRL agent is encouraged to explore moderately hard examples for the inverse dynamics model, but refraining from too difficult ones for the latter to learn.", "The self-regulating feedback structure between the DRL agent and the inverse dynamics model enables them to automatically construct a curriculum for exploration.We perform extensive experiments to validate adversarial exploration strategy on multiple OpenAI gym BID3 robotic arm and hand manipulation task environments simulated by the MuJoCo physics engine (Todorov et al., 2012) , including FetchReach, FetchPush, FetchPickAndPlace, FetchSlide, and HandReach.", "These environments are intentionally selected by us for evaluating the performance of inverse dynamics model, as each of them allows only a very limited set of chained actions to transition the robotic arms and hands to target observations.", "We examine the effectiveness of our method by comparing it against a number of self-supervised IL schemes.", "The experimental results show that our method is more effective and data-efficient than the other self-supervised IL schemes for both low-and high-dimensional observation spaces, as well as in environments with high-dimensional action spaces.", "We also demonstrate that in most of the cases the performance of the inverse dynamics model trained by our method is comparable to that directly trained with expert demonstrations.", "The above observations suggest that our method is superior to the other self-supervised IL schemes even in the absence of human priors.", "We further evaluate our method on environments with action space perturbations, and show that our method is able to achieve satisfactory success rates.", "To justify each of our design decisions, we provide a comprehensive set of ablative analysis and discuss their implications.", "The contributions of this work are summarized as follows:• We introduce an adversarial exploration strategy for self-supervised IL.", "It consists of a DRL agent and an inverse dynamics model developed for efficient exploration and data collection.•", "We employ a competitive scheme for the DRL agent and the inverse dynamics model, enabling them to automatically construct a curriculum for exploration of observation space.•", "We introduce a reward structure for the proposed scheme to stabilize the training process.•", "We demonstrate the proposed method and compare it with a number of baselines for multiple robotic arm and hand manipulation tasks in both low-and high-dimensional state spaces.•", "We validate that our method is generalizable to tasks with high-dimensional action spaces.The remainder of this paper is organized as follows. Section", "2 introduces background material. Section", "3 describes the proposed adversarial exploration strategy in detail. Section", "4 reports the experimental results, and provides an in-depth ablative analysis of our method. Section", "5 concludes.", "In this paper, we presented an adversarial exploration strategy, which consists of a DRL agent and an inverse dynamics model competing with each other for self-supervised IL.", "The former is encouraged to adversarially collect difficult training data for the latter, such that the training efficiency of the latter is significantly enhanced.", "Experimental results demonstrated that our method substantially improved the data collection efficiency in multiple robotic arm and hand manipulation tasks, and boosted the performance of the inverse dynamics model in both low-and high-dimensional observation spaces.", "In addition, we validated that our method is generalizable to environments with high-dimensional action spaces.", "Moreover, we showed that our method is robust to action space perturbations.", "Finally, we provided a set of ablative analysis to validate the effectiveness for each of our design decisions." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.8636363744735718, 0.1538461446762085, 0.05405404791235924, 0, 0.0476190410554409, 0.08695651590824127, 0.0476190410554409, 0.1818181723356247, 0.1111111044883728, 0.0555555522441864, 0.051282044500112534, 0.05882352590560913, 0.0952380895614624, 0.04081632196903229, 0.04651162400841713, 0.13333332538604736, 0.10169491171836853, 0.09756097197532654, 0.11764705181121826, 0.19512194395065308, 0.20408162474632263, 0.1111111044883728, 0.03999999538064003, 0.04878048226237297, 0.0833333283662796, 0.04347825422883034, 0.04999999329447746, 0.06451612710952759, 0, 0.05405404791235924, 0.038461532443761826, 0, 0.0555555522441864, 0.05405404791235924, 0.3384615480899811, 0.07692307233810425, 0.04878048226237297, 0, 0.10810810327529907, 0, 0.04878048226237297, 0, 0.1666666567325592, 0.1428571343421936, 0.026666663587093353, 0.03703703358769417, 0.0555555522441864, 0.07843136787414551, 0.09090908616781235, 0.1463414579629898, 0.04878048226237297, 0.052631575614213943, 0.15789473056793213, 0.15789473056793213, 0.13636362552642822, 0.11764705181121826, 0.04255318641662598, 0.0952380895614624, 0, 0.06666666269302368, 0.11428570747375488, 0.1304347813129425, 0.09999999403953552, 0.07843136787414551, 0.05714285373687744, 0.0624999962747097, 0.05405404791235924 ]

[ "A simple yet effective imitation learning scheme that incentivizes exploration of an environment without any extrinsic reward or human demonstration." ]

[ "This paper proposes a dual variational autoencoder (DualVAE), a framework for generating images corresponding to multiclass labels.", "Recent research on conditional generative models, such as the Conditional VAE, exhibit image transfer by changing labels.", "However, when the dimension of multiclass labels is large, these models cannot change images corresponding to labels, because learning multiple distributions of the corresponding class is necessary to transfer an image.", "This leads to the lack of training data.", "Therefore, instead of conditioning with labels, we condition with latent vectors that include label information.", "DualVAE divides one distribution of the latent space by linear decision boundaries using labels.", "Consequently, DualVAE can easily transfer an image by moving a latent vector toward a decision boundary and is robust to the missing values of multiclass labels.", "To evaluate our proposed method, we introduce a conditional inception score (CIS) for measuring how much an image changes to the target class.", "We evaluate the images transferred by DualVAE using the CIS in CelebA datasets and demonstrate state-of-the-art performance in a multiclass setting.", "Recent conditional generative models have shown remarkable success in generating and transferring images.", "Specifically, a conditional variational autoencoder (CVAE) BID4 can generate conditional images by learning the latent space Z that corresponds to multiclass labels.", "In addition, StarGAN BID1 and FaderNetworks BID5 can generate images corresponding to multiple domains by conditioning with domains such as attributes.However, when the dimension of the multiclass is increased, these models cannot transfer the images corresponding to one arbitrary domain (an element of a multiclass label).", "The possible reasons are the following.", "For simplicity, we consider a binary multiclass classification.", "To transfer an image of a certain class, it is necessary to learn the distributions of the corresponding class.", "That is, assuming that the number of classes in the multiclass is N, conditional models need to create 2 N distributions.", "However, when N is large, training is difficult as O(2 N ) training samples will be required.Hence, instead of conditioning with labels, we propose DualVAE, which conditions with latent vectors that include label information.", "DualVAE divides one distribution of the latent space by N linear decision boundaries which need to learn only O(N ) parameters by adding another decoder p w (y|z) to a variational autoencoder (VAE) BID3 .", "DualVAE assumes that a label is a linear combination of vectors of the latent space and the dual latent space.", "There are two advantages to the DualVAE decoder p w (y|z) being a linear model.", "First, DualVAE can easily transfer an image by moving a latent vector toward a decision boundary.", "Next, DualVAE is robust to the missing values of multiclass labels.In addition to this method, we propose the conditional inception score (CIS), a new metric for conditional transferred images.", "Although the evaluation methods often used in the generation models are the Inception Score (IS) BID9 and the Fréchet Inception Distance BID2 , they are used for evaluating the diversity of images and not suitable for evaluating transferred images conditioned with domains such as attributes or classes.", "Therefore, we propose a new metric to evaluate two properties: the first property pertains to whether images in one domain are transferred properly to images in another domain; the second property pertains to whether images in one domain Figure 1 : Conditional VAE learns 2 n distributions for each binary multiclass label when the number of class is n.", "DualVAE learns n decision boundaries for dividing a distribution of latent space.", "u 1 is a parameter of a decision boundary, which we call a dual vector.transferred to images in another domain can preserve the original properties.", "By using the CIS, we compare DualVAE with other methods that can perform image-to-image translations for multiple domains.In summary, the contributions from this study are as follows:", "1) We introduce DualVAE, a method for transferring images corresponding to multiclass labels and demonstrate that images can be transferred quantitatively and qualitatively.", "2) We propose the CIS, a new metric that can evaluate transferred images corresponding to multiclass labels.", "We proposed DualVAE, a simple framework for generating and transferring images corresponding to multiclass labels.", "Further, we introduced the CIS, a new metric for measuring how much of an image corresponding to the change of labels could be generated.", "The decoder of DualVAE was a simple linear model in this study; however, we would like to test more complex models in the future." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.555555522441864, 0.10810810327529907, 0.43478259444236755, 0.2142857164144516, 0.05882352590560913, 0.29411762952804565, 0.31111109256744385, 0.23255813121795654, 0.25641024112701416, 0.12121211737394333, 0.3414634168148041, 0.3050847351551056, 0.07692307233810425, 0.1428571343421936, 0.37837836146354675, 0.29999998211860657, 0.11764705181121826, 0.19230768084526062, 0.34285715222358704, 0.17142856121063232, 0.05714285373687744, 0.42553192377090454, 0.1428571343421936, 0.375, 0.25, 0.3181818127632141, 0.12765957415103912, 0.3414634168148041, 0.4324324131011963, 0.514285683631897, 0.380952388048172, 0.1860465109348297 ]

[ " a new framework using dual space for generating images corresponding to multiclass labels when the number of class is large" ]

[ "One of the most notable contributions of deep learning is the application of convolutional neural networks (ConvNets) to structured signal classification, and in particular image classification.", "Beyond their impressive performances in supervised learning, the structure of such networks inspired the development of deep filter banks referred to as scattering transforms.", "These transforms apply a cascade of wavelet transforms and complex modulus operators to extract features that are invariant to group operations and stable to deformations.", "Furthermore, ConvNets inspired recent advances in geometric deep learning, which aim to generalize these networks to graph data by applying notions from graph signal processing to learn deep graph filter cascades.", "We further advance these lines of research by proposing a geometric scattering transform using graph wavelets defined in terms of random walks on the graph.", "We demonstrate the utility of features extracted with this designed deep filter bank in graph classification of biochemistry and social network data (incl. state of the art results in the latter case), and in data exploration, where they enable inference of EC exchange preferences in enzyme evolution.", "Over the past decade, numerous examples have established that deep neural networks (i.e., cascades of linear operations and simple nonlinearities) typically outperform traditional \"shallow\" models in various modern machine learning applications, especially given the increasing Big Data availability nowadays.", "Perhaps the most well known example of the advantages of deep networks is in computer vision, where the utilization of 2D convolutions enable network designs that learn cascades of convolutional filters, which have several advantages over fully connected network architectures, both computationally and conceptually.", "Indeed, in terms of supervised learning, convolutional neural networks (ConvNets) hold the current state of the art in image classification, and have become the standard machine learning approach towards processing big structured-signal data, including audio and video processing.", "See, e.g., Goodfellow et al. (2016, Chapter 9) for a detailed discussion.Beyond their performances when applied to specific tasks, pretrained ConvNet layers have been explored as image feature extractors by freezing the first few pretrained convolutional layers and then retraining only the last few layers for specific datasets or applications (e.g., BID47 BID33 .", "Such transfer learning approaches provide evidence that suitably constructed deep filter banks should be able to extract task-agnostic semantic information from structured data, and in some sense mimic the operation of human visual and auditory cortices, thus supporting the neural terminology in deep learning.", "An alternative approach towards such universal feature extraction was presented in BID28 , where a deep filter bank, known as the scattering transform, is designed, rather than trained, based on predetermined families of distruptive patterns that should be eliminated to extract informative representations.", "The scattering transform is constructed as a cascade of linear wavelet transforms and nonlinear complex modulus operations that provides features with guaranteed invariance to a predetermined Lie group of operations such as rotations, translations, or scaling.", "Further, it also provides Lipschitz stability to small diffeomorphisms of the inputted signal.", "Scattering features have been shown to be effective in several audio (e.g., BID6 BID0 BID27 and image (e.g., BID7 BID40 BID34 processing applications, and their advantages over learned features are especially relevant in applications with relatively low data availability, such as quantum chemistry (e.g., BID15 BID35 .Following", "the recent interest in geometric deep learning approaches for processing graph-structured data (see, for example, BID4 and references therein), we present here a generalization of the scattering transform from Euclidean domains to graphs. Similar to", "the Euclidean case, our construction is based on a cascade of bandpass filters, defined in this case using graph signal processing BID38 notions, and complex moduli, which in this case take the form of absolute values (see Sec. 3). While several", "choices of filter banks could generally be used with the proposed cascade, we focus here on graph wavelet filters defined by lazy random walks (see Sec. 2). These wavelet", "filters are also closely related to diffusion geometry and related notions of geometric harmonic analysis, e.g. the diffusion maps algorithm of BID10 and the associated diffusion wavelets of BID11 . Therefore, we", "call the constructed cascade geometric scattering, which also follows the same terminology from geometric deep learning.We note that similar attempts at generalizing the scattering transform to graphs have been presented in BID9 as well as BID49 and BID17 . The latter two", "works are most closely related to the present paper. In them, the authors", "focus on theoretical properties of the proposed graph scattering transforms, and show that such transforms are invariant to graph isomorphism. The geometric scattering", "transform that we define here also possesses the same invariance property, and we expect similar stability properties to hold for the proposed construction as well. However, in this paper we", "focus mainly on the practical applicability of geometric scattering transforms for graph-structured data analysis, with particular emphasis on the task of graph classification, which has received much attention recently in geometric deep learning (see Sec. 4) In supervised graph classification problems one is given a training database of graph/label pairs DISPLAYFORM0 ⊂ G × Y sampled from a set of potential graphs G and potential labels Y. The goal is to use the training data to learn a model f : G → Y that associates to any graph G ∈ G a label y = f (G) ∈ Y. These types of databases arise in biochemistry, in which the graphs may be molecules and the labels some property of the molecule (e.g., its toxicity), as well as in various types of social network databases. Until recently, most approaches", "were kernel based methods, in which the model f was selected from the reproducing kernel Hilbert space generated by a kernel that measures the similarity between two graphs; one of the most successful examples of this approach is the Weisfeiler-Lehman graph kernel of BID37 . Numerous feed forward deep learning", "algorithms, though, have appeared over the last few years. In many of these algorithms, task based", "(i.e., dependent upon the labels Y) graph filters are learned from the training data as part of the larger network architecture. These filters act on a characteristic signal", "x G that is defined on the vertices of any graph G, e.g., x G may be a vector of degrees of each vertex (we remark there are also edge based algorithms, such as BID20 and references within, but these have largely been developed for and tested on databases not considered in Sec. 4). Here, we propose an alternative to these methods", "in the form of a geometric scattering classifier (GSC) that leverages graph-dependent (but not label dependent) scattering transforms to map each graph G to the scattering features extracted from x G . Furthermore, inspired by transfer learning approaches", "such as BID33 , we consider treatment of our scattering cascade as frozen layers on x G , either followed by fully connected classification layers (see FIG2 ), or fed into other classifiers such as SVM or logistic regression. We note that while the formulation in Sec. 3 is phrased", "for a single signal x G , it naturally extends to multiple signals by concatenating their scattering features.In Sec. 4.1 we evaluate the quality of the scattering features and resulting classification by comparing it to numerous graph kernel and deep learning methods over 13 datasets (7 biochemistry ones and 6 social network ones) commonly studied in related literature. In terms of classification accuracy on individual datasets", ", we show that the proposed approach obtains state of the art results on two datasets and performs competitively on the rest, despite only learning a classifier that come after the geometric scattering transform. Furthermore, while other methods may excel on specific datasets", ", when considering average accuracy: within social network data, our proposed GSC outperforms all other methods; in biochemistry or over all datasets, it outperforms nearly all feed forward neural network approaches, and is competitive with state of the art results of graph kernels BID26 and graph recurrent neural networks BID41 . We regard this result as crucial in establishing the universality", "of graph features extracted by geometric scattering, as they provide an effective task-independent representation of analyzed graphs. Finally, to establish their unsupervised qualities, in Sec. 4.2 we", "use geometric scattering features extracted from enzyme data BID2 to infer emergent patterns of enzyme commission (EC) exchange preferences in enzyme evolution, validated with established knowledge from BID12 .", "We presented the geometric scattering transform as a deep filter bank for feature extraction on graphs.", "This transform generalizes the scattering transform, and augments the theoretical foundations of geometric deep learning.", "Further, our evaluation results on graph classification and data exploration show the potential of the produced scattering features to serve as universal representations of graphs.", "Indeed, classification with these features with relatively simple classifier models reaches high accuracy results on most commonly used graph classification datasets, and outperforms both traditional and recent deep learning feed forward methods in terms of average classification accuracy over multiple datasets.", "We note that this might be partially due to the scarcity of labeled big data in this field, compared to more traditional ones (e.g., image or audio classification).", "However, this trend also correlates with empirical results for the classic scattering transform, which excels in cases with low data availability.", "Finally, the geometric scattering features provide a new way for computing and considering global graph representations, independent of specific learning tasks.", "Therefore, they raise the possibility of embedding entire graphs in Euclidean space and computing meaningful distances between graphs with them, which can be used for both supervised and unsupervised learning, as well as exploratory analysis of graph-structured data.APPENDIX A FULL COMPARISON TABLE DISPLAYFORM0 DISPLAYFORM1" ]

[ 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.20408162474632263, 0.1666666567325592, 0.1702127605676651, 0.11538460850715637, 0.36734694242477417, 0.317460298538208, 0.12121211737394333, 0.1269841194152832, 0.13793103396892548, 0.10666666179895401, 0.1230769157409668, 0.2028985470533371, 0.20689654350280762, 0.10256409645080566, 0.08571428060531616, 0.3103448152542114, 0.25806450843811035, 0.18518517911434174, 0.11764705181121826, 0.21875, 0.10526315122842789, 0.25531914830207825, 0.15094339847564697, 0.17741934955120087, 0.24242423474788666, 0.1463414579629898, 0.22641508281230927, 0.1927710771560669, 0.20689654350280762, 0.19999998807907104, 0.2195121943950653, 0.22580644488334656, 0.21052631735801697, 0.11320754140615463, 0.15686273574829102, 0.2857142686843872, 0.25, 0.36734694242477417, 0.25806450843811035, 0.18518517911434174, 0.17391303181648254, 0.3404255211353302, 0.1492537260055542 ]

[ "We present a new feed forward graph ConvNet based on generalizing the wavelet scattering transform of Mallat, and demonstrate its utility in graph classification and data exploration tasks." ]

[ "OCR is inevitably linked to NLP since its final output is in text.", "Advances in document intelligence are driving the need for a unified technology that integrates OCR with various NLP tasks, especially semantic parsing.", "Since OCR and semantic parsing have been studied as separate tasks so far, the datasets for each task on their own are rich, while those for the integrated post-OCR parsing tasks are relatively insufficient.", "In this study, we publish a consolidated dataset for receipt parsing as the first step towards post-OCR parsing tasks.", "The dataset consists of thousands of Indonesian receipts, which contains images and box/text annotations for OCR, and multi-level semantic labels for parsing.", "The proposed dataset can be used to address various OCR and parsing tasks.", "Optical character recognition (OCR) is a technique for converting images of characters into digitized texts [1, 2] .", "Recently, deep learning in computer vision domain has significantly improved the performances of OCR [3, 4] .", "Nonetheless, there is still huge room for improvement, especially concerning the tasks simultaneously linked to natural language processing (NLP) as well.", "In particular, post-OCR parsing is currently one of the most important, yet challenging problems in both OCR and NLP community.", "The goal of post-OCR parsing is to predict pre-defined semantic labels from the given OCR.", "Researchers from both domains have long tried to tackle the problem and collected a significant amount of data sets independently.", "However, since it is a specialized task, the datasets contain critical limitations to provide proper supervision.", "The OCR datasets typically do not have parsing class labels for the extracted texts.", "The parsing datasets usually contain error-free and well-ordered digitized texts in contrast to the erroneous outcomes from OCR process.", "We can add synthetic noise to the parsing data, but the distribution and error patterns could be different from the OCR errors, which would inevitably lead to the degradation of generalization performance.", "Over the past few years, a few post-OCR parsing datasets have been made public through post OCR challenges [5] .", "For example, ICDAR 2019 Post-OCR Challenge introduced the Scanned Receipts OCR and Information Extraction (SROIE) dataset [6] .", "It provides receipt images of texts and two types of annotations for OCR and parsing problem: (1) box-level text annotations for OCR, and (2) document-level parse annotations for parsing.", "Although the availability of both OCR and parsing information have given rise to active research within the field, it still possesses some shortcomings, e.g., limited data size and lack of box-level parsing annotations.", "Considering that only hundreds of samples are provided in the SROIE dataset, weak document-level annotations could not provide enough supervision for training a model with satisfactory performance.", "In this paper, we introduce a novel dataset called CORD, which stands for a Consolidated Receipt Dataset for post-OCR parsing.", "To the best of our knowledge, this is the first publicly available dataset which includes both box-level text and parsing class annotations.", "The parsing class labels are provided in two-levels.", "The eight superclasses include store, payment, menu, subtotal, and total.", "The eight superclasses are subdivided into 54 subclasses e.g., store has nine subclasses including name, address, telephone, and fax.", "Furthermore, it also provides line annotations for the serialization task which is a newly emerging problem as a combination of the two tasks.", "Current semantic parsing techniques can handle only well-ordered texts.", "Texts obtained by OCR, however, are in two-dimensional space, thus we need an appropriate serialization technique for mapping obtained texts into one-dimensional space.", "In our experiments, serialization has a significant impact on parsing performance.", "To recapitulate briefly, the key contributions of our paper are as follows:", "• We introduce a novel and large-scale receipt dataset that can be used for OCR and parsing tasks, from task-specific to end-to-end.", "• Our dataset provides multi-level labels for weakly and strongly supervised parsing tasks.", "The dataset and descriptions will be available on https://github.com/clovaai/cord at the time of publication.", "2 Data Acquisition" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0 ]

[ 0, 0.1875, 0.20512820780277252, 0.5, 0.20689654350280762, 0.260869562625885, 0.14814814925193787, 0, 0.12903225421905518, 0.13333332538604736, 0.1599999964237213, 0.06666666269302368, 0.07692307233810425, 0.1666666567325592, 0.06896550953388214, 0.10526315122842789, 0.2142857164144516, 0.07407406717538834, 0.19354838132858276, 0.04878048226237297, 0.10810810327529907, 0.4285714328289032, 0.12903225421905518, 0.1111111044883728, 0, 0, 0.19354838132858276, 0.10526315122842789, 0.0624999962747097, 0.1904761791229248, 0, 0.5161290168762207, 0.3478260934352875, 0.07999999821186066, 0 ]

[ "We introduce a large-scale receipt dataset for post-OCR parsing tasks." ]

[ "The growth in the complexity of Convolutional Neural Networks (CNNs) is increasing interest in partitioning a network across multiple accelerators during training and pipelining the backpropagation computations over the accelerators.", "Existing approaches avoid or limit the use of stale weights through techniques such as micro-batching or weight stashing.", "These techniques either underutilize of accelerators or increase memory footprint.", "We explore the impact of stale weights on the statistical efficiency and performance in a pipelined backpropagation scheme that maximizes accelerator utilization and keeps memory overhead modest.", "We use 4 CNNs (LeNet-5, AlexNet, VGG and ResNet) and show that when pipelining is limited to early layers in a network, training with stale weights converges and results in models with comparable inference accuracies to those resulting from non-pipelined training on MNIST and CIFAR-10 datasets; a drop in accuracy of 0.4%, 4%, 0.83% and 1.45% for the 4 networks, respectively.", "However, when pipelining is deeper in the network, inference accuracies drop significantly.", "We propose combining pipelined and non-pipelined training in a hybrid scheme to address this drop.", "We demonstrate the implementation and performance of our pipelined backpropagation in PyTorch on 2 GPUs using ResNet, achieving speedups of up to 1.8X over a 1-GPU baseline, with a small drop in inference accuracy.", "Modern Convolutional Neural Networks (CNNs) have grown in size and complexity to demand considerable memory and computational resources, particularly for training.", "This growth makes it sometimes difficult to train an entire network with a single accelerator (Huang et al., 2018; Harlap et al., 2018; .", "Instead, the network is partitioned among multiple accelerators, typically by partitioning its layers among the available accelerators, as shown in Figure 1 for an example 8-layer network.", "The 8 layers are divided into 4 computationally-balanced partitions, P 0 ...P 3 and each partition is mapped to one of the 4 accelerators, A 0 ...A 3 .", "Each accelerator is responsible for the computations associated with the layers mapped to it.", "However, the nature of the backpropagation algorithm used to train CNNs (Rumelhart et al., 1986) is that the computations of a layer are performed only after the computations of the preceding layer in the forward pass of the algorithm and only after the computations of the succeeding layer in the backward pass.", "Further, the computations for one batch of input data are only performed after the computations of the preceding batch have updated the parameters (i.e., weights) of the network.", "These dependences underutilize the accelerators, as shown by the space-time diagram in Figure 2 ; only one accelerator can be active at any given point in time.", "The underutilization of accelerators can be alleviated by pipelining the computations of the backpropagation algorithm over the accelerators (Huang et al., 2018; Harlap et al., 2018; .", "That is, by overlapping the computations of different input data batches using the multiple accelerators.", "However, pipelining causes an accelerator to potentially use weights that are yet to be updated by an accelerator further down in the pipeline.", "The use of such stale weights can negatively affect the statistical efficiency of the network, preventing the convergence of training or producing a model with lower inference accuracy.", "Common wisdom is that the use of stale weights must either be avoided, e.g., with the use of microbatches (Huang et al., 2018) , be constrained to ensure the consistency of the weights within an accelerator using stashing (Harlap et al., 2018) , or by limiting the use of pipelining to very small networks (Mostafa et al., 2017) .", "However, these approaches either underutilize accelerators (Huang et al., 2018) or inflate memory usage to stash multiple copies of weights (Harlap et al., 2018) .", "In this paper we question this common wisdom and explore pipelining that allows for the full utilization of accelerators while using stale weights.", "This results in a pipelining scheme that, compared to existing schemes, is simpler to implement, fully utilizes the accelerators and has lower memory overhead.", "We evaluate this pipelining scheme using 4 CNNs: LeNet-5 (trained on MNIST), AlexNet, VGG and ResNet (all trained on CIFAR-10).", "We analyze the impact of weight staleness and show that if pipelining is limited to early layers in the network, training does converge and the quality of the resulting models is comparable to that of models obtained with non-pipelined training.", "For the 4 networks, the drop in accuracy is 0.4%, 4%, 0.83% and 1.45%, respectively.", "However, inference accuracies drop significantly when the pipelining is deeper in the network.", "While this is not a limitation since the bulk of computations that can benefit from pipelining are in the early convolutional layers, we address this through a hybrid scheme that combines pipelined and non-pipelined training to maintain inference accuracy while still delivering performance improvement.", "Evaluation shows that our pipelined training delivers a speedup of up to 1.8X on a 2-GPU system.", "The remainder of this paper is organized as follows.", "Section 2 briefly describes the backpropagation for training of CNNs.", "Section 3 details our pipelining scheme.", "Section 4 describes how non-pipelined and pipelined backpropagation are combined.", "Section 5 highlights some of the implementation details.", "Experimental evaluation is presented in Section", "6. Related work is reviewed in Section", "7. Finally, Section 8 gives concluding remarks and directions for future work." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.1621621549129486, 0.0714285671710968, 0.0952380895614624, 0.1666666567325592, 0.16129031777381897, 0, 0.1538461446762085, 0.1860465109348297, 0.06451612710952759, 0.1249999925494194, 0, 0.05882352590560913, 0.0833333283662796, 0.0952380895614624, 0.060606054961681366, 0, 0.06451612710952759, 0.07999999821186066, 0, 0.22857142984867096, 0.07692307233810425, 0.060606054961681366, 0.060606054961681366, 0.05882352590560913, 0.06666666269302368, 0.14999999105930328, 0, 0, 0.11764705181121826, 0.2857142686843872, 0.09999999403953552, 0.1904761791229248, 0, 0, 0.10526315122842789, 0, 0, 0 ]

[ "Accelerating CNN training on a Pipeline of Accelerators with Stale Weights" ]

[ "Despite their impressive performance, deep neural networks exhibit striking failures on out-of-distribution inputs.", "One core idea of adversarial example research is to reveal neural network errors under such distribution shifts.", "We decompose these errors into two complementary sources: sensitivity and invariance.", "We show deep networks are not only too sensitive to task-irrelevant changes of their input, as is well-known from epsilon-adversarial examples, but are also too invariant to a wide range of task-relevant changes, thus making vast regions in input space vulnerable to adversarial attacks.", "We show such excessive invariance occurs across various tasks and architecture types.", "On MNIST and ImageNet one can manipulate the class-specific content of almost any image without changing the hidden activations.", "We identify an insufficiency of the standard cross-entropy loss as a reason for these failures.", "Further, we extend this objective based on an information-theoretic analysis so it encourages the model to consider all task-dependent features in its decision.", "This provides the first approach tailored explicitly to overcome excessive invariance and resulting vulnerabilities.", "Figure 1: All images shown cause a competitive ImageNet-trained network to output the exact same probabilities over all 1000 classes (logits shown above each image).", "The leftmost image is from the ImageNet validation set; all other images are constructed such that they match the non-class related information of images taken from other classes (for details see section 2.1).", "The excessive invariance revealed by this set of adversarial examples demonstrates that the logits contain only a small fraction of the information perceptually relevant to humans for discrimination between the classes.Adversarial vulnerability is one of the most iconic failure cases of modern machine learning models BID45 ) and a prime example of their weakness in out-of-distribution generalization.", "It is particularly striking that under i.i.d. settings deep networks show superhuman performance on many tasks BID33 , while tiny targeted shifts of the input distribution can cause them to make unintuitive mistakes.", "The reason for these failures and how they may be avoided or at least mitigated is an active research area BID41 BID20 BID11 .So", "far, the study of adversarial examples has mostly been concerned with the setting of small perturbation, or -adversaries BID23 BID35 BID38 .Perturbation-based", "adversarial examples are appealing because they allow to quantitatively measure notions of adversarial robustness BID9 . However, recent work", "argued that the perturbation-based approach is unrealistically restrictive and called for the need of generalizing the concept of adversarial examples to the unrestricted case, including any input crafted to be misinterpreted by the learned model BID44 BID10 ). Yet, settings beyond", "-robustness are hard to formalize BID19 .We argue here for an", "alternative, complementary viewpoint on the problem of adversarial examples. Instead of focusing", "on transformations erroneously crossing the decision-boundary of classifiers, we focus on excessive invariance as a major cause for adversarial vulnerability. To this end, we introduce", "the concept of invariance-based adversarial examples and show that class-specific content of almost any input can be changed arbitrarily without changing activations of the network, as illustrated in figure 1 for ImageNet. This viewpoint opens up new", "directions to analyze and control crucial aspects underlying vulnerability to unrestricted adversarial examples.The invariance perspective suggests that adversarial vulnerability is a consequence of narrow learning, yielding classifiers that rely only on few highly predictive features in their decisions. This has also been supported", "by the observation that deep networks strongly rely on spectral statistical regularities BID29 , or stationary statistics BID17 to make their decisions, rather than more abstract features like shape and appearance. We hypothesize that a major", "reason for this excessive invariance can be understood from an information-theoretic viewpoint of crossentropy, which maximizes a bound on the mutual information between labels and representation, giving no incentive to explain all class-dependent aspects of the input. This may be desirable in some", "cases, but to achieve truly general understanding of a scene or an object, machine learning models have to learn to successfully separate essence from nuisance and subsequently generalize even under shifted input distributions.", "Failures of deep networks under distribution shift and their difficulty in out-of-distribution generalization are prime examples of the limitations in current machine learning models.", "The field of adversarial example research aims to close this gap from a robustness point of view.", "While a lot of work has studied -adversarial examples, recent trends extend the efforts towards the unrestricted case.", "However, adversarial examples with no restriction are hard to formalize beyond testing error.", "We introduce a reverse view on the problem to: (1) show that a major cause for adversarial vulnerability is excessive invariance to semantically meaningful variations, (2) demonstrate that this issue persists across tasks and architectures; and (3) make the control of invariance tractable via fully-invertible networks.In summary, we demonstrated how a bijective network architecture enables us to identify large adversarial subspaces on multiple datasets like the adversarial spheres, MNIST and ImageNet.", "Afterwards, we formalized the distribution shifts causing such undesirable behavior via information theory.", "Using this framework, we find one of the major reasons is the insufficiency of the vanilla cross-entropy loss to learn semantic representations that capture all task-dependent variations in the input.", "We extend the loss function by components that explicitly encourage a split between semantically meaningful and nuisance features.", "Finally, we empirically show that this split can remove unwanted invariances by performing a set of targeted invariance-based distribution shift experiments." ]

[ 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.1304347813129425, 0.11999999731779099, 0.045454543083906174, 0.9166666865348816, 0.08888888359069824, 0.039215680211782455, 0.1249999925494194, 0.0714285671710968, 0.04255318641662598, 0.07017543166875839, 0.0634920597076416, 0.16867469251155853, 0.17910447716712952, 0, 0.07547169178724289, 0.1599999964237213, 0.11764705181121826, 0.13636362552642822, 0.09090908616781235, 0.1071428507566452, 0.1492537260055542, 0.21917808055877686, 0.1764705777168274, 0.13698630034923553, 0.15625, 0.2181818187236786, 0.16326530277729034, 0.07999999821186066, 0.1304347813129425, 0.1505376249551773, 0, 0.1355932205915451, 0.07843136787414551, 0.1111111044883728 ]

[ "We show deep networks are not only too sensitive to task-irrelevant changes of their input, but also too invariant to a wide range of task-relevant changes, thus making vast regions in input space vulnerable to adversarial attacks." ]

[ "Flow-based generative models are powerful exact likelihood models with efficient sampling and inference. \n", "Despite their computational efficiency, flow-based models generally have much worse density modeling performance compared to state-of-the-art autoregressive models.", "In this paper, we investigate and improve upon three limiting design choices employed by flow-based models in prior work: the use of uniform noise for dequantization, the use of inexpressive affine flows, and the use of purely convolutional conditioning networks in coupling layers.", "Based on our findings, we propose Flow++, a new flow-based model that is now the state-of-the-art non-autoregressive model for unconditional density estimation on standard image benchmarks.", "Our work has begun to close the significant performance gap that has so far existed between autoregressive models and flow-based models.", "Deep generative models -latent variable models in the form of variational autoencoders BID16 , implicit generative models in the form of GANs BID8 , and exact likelihood models like PixelRNN/CNN (van den c) , Image Transformer BID22 , PixelSNAIL , NICE, RealNVP, and Glow BID5 BID15 -have recently begun to successfully model high dimensional raw observations from complex real-world datasets, from natural images and videos, to audio signals and natural language BID14 BID34 .Autoregressive", "models, a certain subclass of exact likelihood models, achieve state-of-the-art density estimation performance on many challenging real-world datasets, but generally suffer from slow sampling time due to their autoregressive structure BID28 BID22 . Inverse autoregressive", "models can sample quickly and potentially have strong modeling capacity, but they cannot be trained efficiently by maximum likelihood . Non-autoregressive flow-based", "models (which we will refer to as \"flow models\"), such as NICE, RealNVP, and Glow, are efficient for sampling, but have so far lagged behind autoregressive models in density estimation benchmarks BID5 BID15 .In the hope of creating an ideal", "likelihood-based generative model that simultaneously has fast sampling, fast inference, and strong density estimation performance, we seek to close the density estimation performance gap between flow models and autoregressive models. In subsequent sections, we present", "our new flow model, Flow++, which is powered by an improved training procedure for continuous likelihood models and a number of architectural extensions of the coupling layer defined by BID5 .", "We presented Flow++, a new flow-based generative model that begins to close the performance gap between flow models and autoregressive models.", "Our work considers specific instantiations of design principles for flow models -dequantization, flow design, and conditioning architecture design -and we hope these principles will help guide future research in flow models and likelihoodbased models in general.7", "APPENDIX A: SAMPLES" ]

[ 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.2222222238779068, 0.19354838132858276, 0.16326530277729034, 0.2631579041481018, 0.1818181723356247, 0.11428570747375488, 0.17391303181648254, 0.17142856121063232, 0.23076923191547394, 0.23255813121795654, 0.1904761791229248, 0.23529411852359772, 0.1395348757505417, 0 ]

[ "Improved training of current flow-based generative models (Glow and RealNVP) on density estimation benchmarks" ]

[ "Modern deep artificial neural networks have achieved impressive results through models with orders of magnitude more parameters than training examples which control overfitting with the help of regularization.", "Regularization can be implicit, as is the case of stochastic gradient descent and parameter sharing in convolutional layers, or explicit.", "Explicit regularization techniques, most common forms are weight decay and dropout, have proven successful in terms of improved generalization, but they blindly reduce the effective capacity of the model, introduce sensitive hyper-parameters and require deeper and wider architectures to compensate for the reduced capacity.", "In contrast, data augmentation techniques exploit domain knowledge to increase the number of training examples and improve generalization without reducing the effective capacity and without introducing model-dependent parameters, since it is applied on the training data.", "In this paper we systematically contrast data augmentation and explicit regularization on three popular architectures and three data sets.", "Our results demonstrate that data augmentation alone can achieve the same performance or higher as regularized models and exhibits much higher adaptability to changes in the architecture and the amount of training data.", "One of the central issues in machine learning research and application is finding ways of improving generalization.", "Regularization, loosely defined as any modification applied to a learning algorithm that helps prevent overfitting, plays therefore a key role in machine learning (Girosi et al., 1995; Müller, 2012) .", "In the case of deep learning, where neural networks tend to have several orders of magnitude more parameters than training examples, statistical learning theory (Vapnik & Chervonenkis, 1971) indicates that regularization becomes even more crucial.", "Accordingly, a myriad of techniques have been proposed as regularizers: weight decay (Hanson & Pratt, 1989) and other L p penalties; dropout (Srivastava et al., 2014) and stochastic depth (Huang et al., 2016) , to name a few examples.", "Moreover, whereas in simpler machine learning algorithms the regularizers can be easily identified as explicit terms in the objective function, in modern deep neural networks the sources of regularization are not only explicit, but implicit (Neyshabur et al., 2014) .", "In this regard, many techniques have been studied for their regularization effect, despite not being explicitly intended as such.", "That is the case of unsupervised pre-training (Erhan et al., 2010) , multi-task learning (Caruana, 1998) , convolutional layers (LeCun et al., 1990) , batch normalization (Ioffe & Szegedy, 2015) or adversarial training (Szegedy et al., 2013) .", "In sum, there are multiple elements in deep learning that contribute to reduce overfitting and thus improve generalization.", "Driven by the success of such techniques and the efficient use of GPUs, considerable research effort has been devoted to finding ways of training deeper and wider networks with larger capacity (Simonyan & Zisserman, 2014; He et al., 2016; Zagoruyko & Komodakis, 2016) .", "Ironically, the increased representational capacity is eventually reduced in practice by the use of explicit regularization, most commonly weight decay and dropout.", "It is known, for instance, that the gain in generalization provided by dropout comes at the cost of using larger models and training for longer (Goodfellow et al., 2016) .", "Hence, it seems that with these standard regularization methods deep networks are wasting capacity (Dauphin & Bengio, 2013) .", "Unlike explicit regularization, data augmentation improves generalization without reducing the capacity of the model.", "Data augmentation, that is synthetically expanding a data set by apply-ing transformations on the available examples, has been long used in machine learning (Simard et al., 1992) and identified as a critical component of many recent successful models, like AlexNet (Krizhevsky et al., 2012) , All-CNN (Springenberg et al., 2014) or ResNet (He et al., 2016) , among others.", "Although it is most popular in computer vision, data augmentation has also proven effective in speech recognition (Jaitly & Hinton, 2013) , music source separation (Uhlich et al., 2017) or text categorization (Lu et al., 2006) .", "Today, data augmentation is an almost ubiquitous technique in deep learning, which can also be regarded as an implicit regularizer for it improves generalization.", "Recently, the deep learning community has become more aware of the importance of data augmentation (Hernández-García & König, 2018b) and new techniques, such as cutout (DeVries & Taylor, 2017a) or augmentation in the feature space (DeVries & Taylor, 2017b) , have been proposed.", "Very interestingly, a promising avenue for future research has been set by recently proposed models that automatically learn the data transformations (Hauberg et al., 2016; Lemley et al., 2017; Ratner et al., 2017; Antoniou et al., 2017) .", "Nonetheless, another study by Perez & Wang (2017) analyzed the performance of different techniques for object recognition and concluded that one of the most successful techniques so far is still the traditional data augmentation carried out in most studies.", "However, despite its popularity, the literature lacks, to our knowledge, a systematic analysis of the impact of data augmentation on convolutional neural networks compared to explicit regularization.", "It is a common practice to train the models with both explicit regularization, typically weight decay and dropout, and data augmentation, assuming they all complement each other.", "Zhang et al. (2017) included data augmentation in their analysis of generalization of deep networks, but it was questionably considered an explicit regularizer similar to weight decay and dropout.", "To our knowledge, the first time data augmentation and explicit regularization were systematically contrasted was the preliminary study by Hernández-García & König (2018b) .", "The present work aims at largely extending that work both with more empirical results and a theoretical discussion.", "Our specific contributions are the following:", "• Propose definitions of explicit and implicit regularization that aim at solving the ambiguity in the literature (Section 2).", "• A theoretical discussion based on statistical learning theory about the differences between explicit regularization and data augmentation, highlighting the advantages of the latter (Section 3).", "• An empirical analysis of the performance of models trained with and without explicit regularization, and different levels of data augmentation on several benchmarks (Sections 4 and 5).", "Further, we study their adaptability to learning from fewer examples (Section 5.2) and to changes in the architecture (Section 5.3).", "• A discussion on why encouraging data augmentation instead of explicit regularization can benefit both theory and practice in deep learning (Section 6).", "2 EXPLICIT AND IMPLICIT REGULARIZATION Zhang et al. (2017) raised the thought-provoking idea that \"explicit regularization may improve generalization performance, but is neither necessary nor by itself sufficient for controlling generalization error.\"", "The authors came to this conclusion from the observation that turning off the explicit regularizers of a model does not prevent the model from generalizing reasonably well.", "This contrasts with traditional machine learning involving convex optimization, where regularization is necessary to avoid overfitting and generalize (Vapnik & Chervonenkis, 1971) .", "Such observation led the authors to suggest the need for \"rethinking generalization\" in order to understand deep learning.", "We argue it is not necessary to rethink generalization if we instead rethink regularization and, in particular, data augmentation.", "Despite their thorough analysis and relevant conclusions, Zhang et al. (2017) arguably underestimated the role of implicit regularization and considered data augmentation an explicit form of regularization much like weight decay and dropout.", "This illustrates that the terms explicit and implicit regularization have been used subjectively and inconsistently in the literature before.", "In order to avoid the ambiguity and facilitate the discussion, we propose the following definitions of explicit and implicit regularization 1 :", "• Explicit regularization techniques are those which reduce the representational capacity of the model they are applied on.", "That is, given a model class H 0 , for instance a neural network architecture, the introduction of explicit regularization will span a new hypothesis set H 1 , which is a proper subset of the original set, i.e. H 1 H 0 .", "• Implicit regularization is the reduction of the generalization error or overfitting provided by means other than explicit regularization techniques.", "Elements that provide implicit regularization do not reduce the representational capacity, but may affect the effective capacity of the model, that is the achievable set of hypotheses given the model, the optimization algorithm, hyperparameters, etc.", "One of the most common explicit regularization techniques in machine learning is L p -norm regularization, of which weight decay is a particular case, widely used in deep learning.", "Weight decay sets a penalty on the L 2 norm of the learnable parameters, thus constraining the representational capacity of the model.", "Dropout is another common example of explicit regularization, where the hypothesis set is reduced by stochastically deactivating a number of neurons during training.", "Similar to dropout, stochastic depth, which drops whole layers instead of neurons, is also an explicit regularization technique.", "There are multiple elements in deep neural networks that implicitly regularize the models.", "Note, in this regard, that the above definition, contrary to explicit regularization, does not refer to techniques, but to a regularization effect, as it can be provided by elements of very different nature.", "For instance, stochastic gradient descent (SGD) is known to have an implicit regularization effect without constraining the representational capacity.", "Batch normalization does not either reduce the capacity, but it improves generalization by smoothing the optimization landscape Santurkar et al. (2018) .", "Of quite a different nature, but still implicit, is the regularization effect provided by early stopping, which does not reduce the representational, but the effective capacity.", "By analyzing the literature, we identified some previous pieces of work which, lacking a definition of explicit and implicit regularization, made a distinction apparently based on the mere intention of the practitioner.", "Under such notion, data augmentation has been considered in some cases an explicit regularization technique, as in Zhang et al. (2017) .", "Here, we have provided definitions for explicit and implicit regularization based on their effect on the representational capacity and argue that data augmentation is not explicit, but implicit regularization, since it does not affect the representational capacity of the model.", "We have presented a systematic analysis of the role of data augmentation in deep convolutional neural networks for object recognition, focusing on the comparison with popular explicit regularization techniques-weight decay and dropout.", "In order to facilitate the discussion and the analysis, we first proposed in Section 2 definitions of explicit and implicit regularization, which have been ambiguously used in the literature.", "Accordingly, we have argued that data augmentation should not be considered an explicit regularizer, such as weight decay and dropout.", "Then, we provided some theoretical insights in Section 3 that highlight some advantages of data augmentation over explicit regularization.", "Finally, we have empirically shown that explicit regularization is not only unnecessary (Zhang et al., 2017) , but also that its generalization gain can be achieved by data augmentation alone.", "Moreover, we have demonstrated that, unlike data augmentation, weight decay and dropout exhibit poor adaptability to changes in the architecture and the amount of training data.", "Despite the limitations of our empirical study, we have chosen three significantly distinct network architectures and three data sets in order to increase the generality of our conclusions, which should ideally be confirmed by future work on a wider range of models, data sets and even other domains such text or speech.", "It is important to note, however, that we have taken a conservative approach in our experimentation: all the hyperparameters have been kept as in the original models, which included both weight decay and dropout, as well as light augmentation.", "This setup is clearly suboptimal for models trained without explicit regularization.", "Besides, the heavier data augmentation scheme was deliberately not optimized to improve the performance and it was not the scope of this work to propose a specific data augmentation technique.", "As future work, we plan to propose data augmentation schemes that can more successfully be exploited by any deep model.", "The relevance of our findings lies in the fact that explicit regularization is currently the standard tool to enable the generalization of most machine learning methods and is included in most convolutional neural networks.", "However, we have empirically shown that simply removing the explicit regularizers often improves the performance or only marginally reduces it, if some data augmentation is applied.", "These results are supported by the theoretical insights provided in in Section 3.", "Zhang et al. (2017) suggested that regularization might play a different role in deep learning, not fully explained by statistical learning theory (Vapnik & Chervonenkis, 1971) .", "We have argued instead that the theory still naturally holds in deep learning, as long as one considers the crucial role of implicit regularization: explicit regularization seems to be no longer necessary because its contribution is already provided by the many elements that implicitly and successfully regularize the models: to name a few, stochastic gradient descent, convolutional layers and data augmentation." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.24561403691768646, 0.23529411852359772, 0.260869562625885, 0.2295081913471222, 0.21276594698429108, 0.4067796468734741, 0.1702127605676651, 0.1355932205915451, 0.21875, 0.21212120354175568, 0.2686567008495331, 0.11999999731779099, 0.09677419066429138, 0.12244897335767746, 0.19999998807907104, 0.26923075318336487, 0.16949151456356049, 0.12244897335767746, 0.22727271914482117, 0.1463414579629898, 0.0937499925494194, 0.14814814925193787, 0.21212120354175568, 0.06451612710952759, 0.1846153736114502, 0.3272727131843567, 0.31578946113586426, 0.3050847351551056, 0.22641508281230927, 0.0833333283662796, 0.054054051637649536, 0.2448979616165161, 0.2181818187236786, 0.290909081697464, 0.23999999463558197, 0.25925925374031067, 0.0634920597076416, 0.18518517911434174, 0.15094339847564697, 0.12765957415103912, 0.2448979616165161, 0.29999998211860657, 0.2083333283662796, 0.23999999463558197, 0.12765957415103912, 0.15625, 0.20408162474632263, 0.17241378128528595, 0.25, 0.12244897335767746, 0.1538461446762085, 0.16326530277729034, 0.1818181723356247, 0.25806450843811035, 0.11999999731779099, 0.07843136787414551, 0.1111111044883728, 0.13793103396892548, 0.23529411852359772, 0.2539682388305664, 0.39344263076782227, 0.2142857164144516, 0.31372547149658203, 0.2448979616165161, 0.1666666567325592, 0.4814814627170563, 0.18666666746139526, 0.2461538463830948, 0.1428571343421936, 0.25925925374031067, 0.15686273574829102, 0.3050847351551056, 0.1428571343421936, 0.09302325546741486, 0.10526315122842789, 0.23529411852359772 ]

[ "Deep neural networks trained with data augmentation do not require any other explicit regularization (such as weight decay and dropout) and exhibit greater adaptaibility to changes in the architecture and the amount of training data." ]

[ "Adversarial feature learning (AFL) is one of the promising ways for explicitly constrains neural networks to learn desired representations; for example, AFL could help to learn anonymized representations so as to avoid privacy issues.", "AFL learn such a representations by training the networks to deceive the adversary that predict the sensitive information from the network, and therefore, the success of the AFL heavily relies on the choice of the adversary.", "This paper proposes a novel design of the adversary, {\\em multiple adversaries over random subspaces} (MARS) that instantiate the concept of the {\\em volunerableness}.", "The proposed method is motivated by an assumption that deceiving an adversary could fail to give meaningful information if the adversary is easily fooled, and adversary rely on single classifier suffer from this issues. \n", "In contrast, the proposed method is designed to be less vulnerable, by utilizing the ensemble of independent classifiers where each classifier tries to predict sensitive variables from a different {\\em subset} of the representations. \n", "The empirical validations on three user-anonymization tasks show that our proposed method achieves state-of-the-art performances in all three datasets without significantly harming the utility of data. \n", "This is significant because it gives new implications about designing the adversary, which is important to improve the performance of AFL.", "Since its invention over ten years ago BID4 , deep neural networks (DNN) have shown significant performance improvements in various fields.", "When we apply DNN or more general machine learning techniques to real-world data, one of the key challenges is how to systematically incorporate the desired constraints into the learned representations in a controllable manner.", "For example, when practitioners apply these techniques to the data that contain a lot of user information (such as images with username BID1 or data of wearables BID6 ), the desired representations should not contain user-information that may result in privacy issues.", "Moreover, for legal and ethical reasons, machine learning algorithms have to make fair decisions, which do not rely on sensitive variables such as gender, age, or race BID8 BID1 .", "Such a background requires removal of information related to specific factors (such as user ID, race, etc.) from the representation; this is called censoring representations in this paper.One of the recently proposed approaches for censoring representation is adversarial feature learning (AFL) BID1 BID6 BID13 , which employs the adversarial training framework to constrain the representations FIG0 .", "Specifically, AFL considers an adversarial classifier who attempts to predict sensitive variables from the representations of a DNN and simultaneously trains the DNN to deceive the classifier.", "By alternatively or jointly (using gradient reversal layer proposed by BID2 ) training the adversary and DNN in such a manner, AFL ensures that there is little or no information about the sensitive variables in the representations.Although some previous studies report significant performance improvements of the AFL in the context of censoring representations, the success of the AFL depends on the choice of the adversarial classifier.", "For example, if we use a logistic regression as the adversarial classifier, AFL can only eliminate the information that is linearly separated in the representation spaces and cannot remove any non-linear dependency.", "It is also possible that deceiving some classifier might be too easy, resulting in poor performance improvements of AFL.", "As such, the design of adversary is crucial for the performance of AFL; however, existing studies fail to address how to design the adversary for improving the quality of AFL.In this paper, we propose a novel design of adversary for improving the performance of AFL, multiple-adversaries over random subspace (MARS), which consider the vulnerableness of the adversary.", "The proposed design is motivated by the recent report BID6 that is just increasing the capacity of adversary did not successfully improves the performance of AFL BID6 , and assumptions that deceiving an adversary fail to give meaningful information if the adversary is easily fooled, and adversary relies on single classifier suffer from this issues.", "The proposed method incorporates multiple adversaries where each adversary tries to predict sensitive variables from a different subset of the representations.", "This design makes adversary less vulnerable to the update of the encoder since the encoder needs to in a set of diverse adversaries.", "In this paper, we validate the effectiveness of the proposed design by empirically showing that (1) MARS archives better performance compared to baselines (that uses a single adversary and multiple adversaries over the entire representation spaces), and (2) MARS is less vulnerable compared to the baselines.The primary contributions of this paper are as follows:• This is the first study verifying the importance of the design of adversary in AFL and proposes the novel design for improving AFL.", "This is significant because the results suggest that design of adversary is vital for the performance of adversary, and gives new implications about designing the adversary in AFL, which is important to improve the performance of AFL.", "It is worth mentioning that, except our paper, all existing studies focus only on the accuracy/capacity for designing adversaries, which is not enough for improving the performance of AFL as shown in this paper.•", "The proposed method achieved state-of-the-art performance in the task of censoring representations, which is essential to extend the applicability of DNN to many real-world applications. The", "empirical validation using three user-anonymization tasks shows that the proposed method allows the learning of significantly more anonymized representations with negligible performance degradation. Specifically", ", the probability of correctly predicting the user ID from learned representations is more than 0.07 points better on average than that of a single adversary and multiple adversaries over entire representation spaces.2 PROBLEM DEFINITION", "AND RELATED WORKS 2.1 PROBLEM DEFINITION: CENSORING REPRESENTATIONS Censoring representation is a task to obtaining unbiased features. Here, unbiased features", "are features that are less affected by S, where S is a random variable that we want to remove from the data for some reason. One typical reason is related", "to fairness or privacy, which requires the output of neural networks not to be affected by unfair information or not contain user information.It should be noted that poor design of the censoring procedure significantly reduces the utility of data. For example, the output of random", "mapping f rand apparently has no information about S, but it also gives no information about target Y . Alternatively, as a more realistic", "example, a neural network with limited capacity possibly acquires less information about S, but it may also result in poorer performance. Therefore, the primary goal of censoring", "representation is to obtain an encoder E that reduces information about S, while maintaining information about Y . Formally, the task can be written as a joint", "optimization problem of the loss: DISPLAYFORM0 where X indicates the input random variable, E is an encoder that transforms X to representation R, λ is the weighting parameter, and V and L are loss functions that represent how much information about S and Y is present, respectively. Note that S can be any form of variables such", "as binary variable, categorical variable, or continuous variable. In this paper, we primarily consider a particular", "variant of censoring representation tasks, where we learn E with deep neural networks and S is the user ID (anonymization tasks).", "This study proposed MARS, which incorporates multiple adversaries where each adversary has a different role and conducted empirical validations on the efficacy of the proposed method for censoring representations, specifically user-anonymization for the data of wearables.", "TAB0 compares the proposed method and several baselines and shows the efficacy of the proposed method against various evaluators.", "Figure 2 qualitatively shows that the proposed method provides wellanonymized representations.", "FIG4 -c shows that each adversary in MARS has the diverse role, resulting MARS more robust to the update of E as a whole.", "All these results support that the proposed method is more effective in removing the influence of a specific factor (user in experiments) compared to the previous methods.One of the reasons why MARS works well is that the adversary is designed to have diverse-views by incorporating random subspace methods, resulting the encoder need to be stronger to deceive the adversary.", "It is worth mentioning that the capacity or accuracy of the adversary is not the only a definitive factor that determines the success of the adversarial feature learning, as shown by the superior performance of MARS over MA that has 1 1−α times the larger capacity of MARS.", "Moreover, the final performance of AFL is significantly different even if the accuracy of D is reasonably similar during training, as shown in FIG4 -b.", "As mentioned in the related work section, such knowledge is essential to design the adversary in practice, and prior studies of adversarial feature learning did not address this issues.Although this paper focused on the case where the subsets are randomly selected and fixed, this might not be necessary.", "One of the possible extensions is to determine subsets with more sophisticated ways (e.g., by performing clustering or soft-clustering on the representation spaces after few training iterations), or to learn how to select the subset itself by adding the criterion regarding the diversity of adversaries.", "Also, it might be possible to realize the diversity of adversaries by methods other than subspace selection.", "One possible way is to constrain weights of two adversaries so that they are an orthogonal view, which is used in semi-supervised learning using co-training BID11 , or it might be worth a try to add different noises for each adversary.It might be worth mentioning about applicability and implications of MARS for other applications of adversarial training, such as image generation.", "From the perspective of the applicability, the MARS itself does not rely on any domain-specific settings and is therefore general enough for many applications based on adversarial training.", "For example, we can build multiple-adversaries upon the subset of feature spaces (maybe not on the image spaces).", "This makes discriminator have diverse-view, so it might be useful for preventing mode collapse that is one of the well-known problems in imagegeneration with adversarial training.", "In the context of image-generation, Generative Multi Adversarial Networks proposed by BID0 , which also use multiple adversaries, shows that multiple adversaries are useful for generating better images, and for avoiding mode collapse.", "It might be interesting to see if enhancing the diversity of discriminators by preparing asymmetric adversaries as with this paper helps to generate a better image or to avoid mode collapse better.", "Table 2 shows the selected λ for each combination of datasets and baselines.", "Although the best hyper-parameter might be determined by the balance between log q M and log q D , here we cannot see the obvious relationships between the best λ and the easiness of tasks." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.25, 0.19230768084526062, 0.260869562625885, 0.21052631735801697, 0.24561403691768646, 0.1538461446762085, 0.17777776718139648, 0, 0.10526315122842789, 0.0952380895614624, 0.072727270424366, 0.2702702581882477, 0.1666666567325592, 0.1818181723356247, 0.0714285671710968, 0.04444443807005882, 0.1904761791229248, 0.2028985470533371, 0.21276594698429108, 0.22727271914482117, 0.2195121943950653, 0.2222222238779068, 0.13793103396892548, 0.2083333283662796, 0.12244897335767746, 0.09999999403953552, 0.04444443807005882, 0.1538461446762085, 0.13333332538604736, 0, 0.07547169178724289, 0.07999999821186066, 0.0833333283662796, 0, 0.08510638028383255, 0.24561403691768646, 0.14999999105930328, 0.10810810327529907, 0.1666666567325592, 0.19718308746814728, 0.19999998807907104, 0.0833333283662796, 0.20895521342754364, 0.1249999925494194, 0.1860465109348297, 0.1249999925494194, 0.15686273574829102, 0.1395348757505417, 0.19230768084526062, 0.17543859779834747, 0.1818181723356247, 0.1538461446762085, 0.11538460850715637 ]

[ "This paper improves the quality of the recently proposed adversarial feature leaning (AFL) approach for incorporating explicit constrains to representations, by introducing the concept of the {\\em vulnerableness} of the adversary. " ]

[ "A key problem in neuroscience, and life sciences more generally, is that data is generated by a hierarchy of dynamical systems.", "One example of this is in \\textit{in-vivo} calcium imaging data, where data is generated by a lower-order dynamical system governing calcium flux in neurons, which itself is driven by a higher-order dynamical system of neural computation.", "Ideally, life scientists would be able to infer the dynamics of both the lower-order systems and the higher-order systems, but this is difficult in high-dimensional regimes.", "A recent approach using sequential variational auto-encoders demonstrated it was possible to learn the latent dynamics of a single dynamical system for computations during reaching behaviour in the brain, using spiking data modelled as a Poisson process.", "Here we extend this approach using a ladder method to infer a hierarchy of dynamical systems, allowing us to capture calcium dynamics as well as neural computation.", "In this approach, spiking events drive lower-order calcium dynamics, and are themselves controlled by a higher-order latent dynamical system.", "We generate synthetic data by generating firing rates, sampling spike trains, and converting spike trains to fluorescence transients, from two dynamical systems that have been used as key benchmarks in recent literature: a Lorenz attractor, and a chaotic recurrent neural network.", "We show that our model is better able to reconstruct Lorenz dynamics from fluorescence data than competing methods.", "However, though our model can reconstruct underlying spike rates and calcium transients from the chaotic neural network well, it does not perform as well at reconstructing firing rates as basic techniques for inferring spikes from calcium data.", "These results demonstrate that VLAEs are a promising approach for modelling hierarchical dynamical systems data in the life sciences, but that inferring the dynamics of lower-order systems can potentially be better achieved with simpler methods.", "Many datasets in the life sciences are generated by a hierarchy of dynamical systems, wherein lower-order dynamical systems that directly generate the data are driven by higher-order dynamical systems that are not observable.", "This problem is outlined in figure 1A , in which noisy observations x depend on the state z 1 of a low-order dynamical system that is perturbed by inputs u 1 .", "The state of this dynamical system is also coupled to the state z 2 of a higher-order dynamical system, which can be perturbed independently by inputs u 2 .", "One example of such a system in in-vivo two-photon calcium imaging from neuroscience.", "Calcium imaging provides systems neuroscientists with the ability to observe the activity of hundreds of neurons simultaneously during behavioural experiments.", "Such experiments have allowed neuroscientists to ask questions about the underlying computations and algorithms that neural circuits are implementing in perception, decision-making, memory, and many other processes.", "Such experiments can be characterized as observing a hierarchical dynamical system (Fig 1B) in which measurable calcium fluorescence is primarily determined by dynamics based on voltage-gated calcium channels and calcium binding to fluorescence dyes, and the rate of fluorescence transients controlled by the underlying computation.", "Recent applications of sequential variational autoencoders to neural data analysis has seen great success in inferring underlying computations in populations of cells in macaque and human motor cortex Pandarinath et al. (2018) .", "By characterizing neural computation as low-dimensional dynamic factors in a non-hierarchical dynamical systems, Pandarinath et al. (2018) showed that these dynamic factors trained to generate the inhomogeneous intensity functions explaining the rate of spikes assumed to follow a Poisson process.", "Crucially, these low-dimensional factors could also decode reaching behaviour of macaques and humans with much higher fidelity than any other dimensionality reduction method.", "Although this is a significant advance in our ability to analyze neural data in the form of spikes trains, two-photon calcium imaging poses the additional problem of identifying latent spike trains in fluorescence traces.", "This problem has been independently addressed in a number of different ways, including deconvolution Friedrich et al. (2017) and variational inference Speiser et al. (2017) .", "If we continue to model the frequency of events as being generated by a Poisson process, this can be seen as hierarchy of dynamical systems (Fig 1A) , in which low dimensional dynamics generate spike probabilities that in turn drive fluctuations in biophysical dynamics of calcium activity ( Fig 1B.", "Here we propose a method that extends LFADS to accommodate calcium activity using this hierarchical dynamical systems approach, in which we can infer both the latent dynamics and the latent spike trains from the observed calcium fluorescence signal.", "We present a hierarchical recurrent variational autoencoder model capable of reconstructing latent dynamics, latent spike trains, and calcium fluorescence traces in a benchmark synthetic dataset.", "Of the four methods tested, our model is the only one capable of reconstructing all three.", "Furthermore, our model performed best in reconstructing latent dynamics in our synthetic dataset We will need to assess our model on further synthetic benchmark data to assess the validity of our approach.", "Since our model is trained end-to-end, it should be possible to extend to reconstructing raw 2-photon imaging videos, which could enable us to train models to uncover latent dynamics from arbitrarily shaped neuronal structures.", "This would of great use to neuroscientists who are largely restricted to techniques that extract fluorescence traces from regions of interest with somatic shapes, whereas the morphological diversity of dendrites is much greater.", "An additional advantage of using our hierarchical model is that we can obtain measures of the uncertainty in both the latent dynamics, and the latent spike trains.", "The correlation in uncertainty between layers of this hierarchy may be what allows superior inference of the latent dynamics, despite less accurate reconstructions of the spike trains than OASIS, which provides no measure of uncertainty.", "We hope to improve our model to better capture the relationships between layers of this hierarchy in future.", "We describe a use-case in neuroscience (2-photon calcium imaging data) for which this model may be very useful.", "However, we are keen to investigate the general case of hierarchical dynamical systems and their utility in uncovering structure in datasets outside this domain.", "A APPENDIX" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0 ]

[ 0.2380952388048172, 0.12244897335767746, 0.21739129722118378, 0.2857142686843872, 0.3478260934352875, 0.04878048226237297, 0.19999998807907104, 0.14999999105930328, 0.1090909019112587, 0.2222222238779068, 0.25531914830207825, 0.1599999964237213, 0.17391303181648254, 0.22857142984867096, 0.19999998807907104, 0.1249999925494194, 0.1666666567325592, 0.15686273574829102, 0.21052631735801697, 0.08888888359069824, 0.19230768084526062, 0.1818181723356247, 0.24242423474788666, 0.2545454502105713, 0.35555556416511536, 0.1621621549129486, 0.260869562625885, 0.11320754140615463, 0.11538460850715637, 0.2222222238779068, 0.1538461446762085, 0.3589743673801422, 0.29999998211860657, 0.2222222238779068 ]

[ "We extend a successful recurrent variational autoencoder for dynamic systems to model an instance of dynamic systems hierarchy in neuroscience using the ladder method." ]

[ "With the proliferation of specialized neural network processors that operate on low-precision integers, the performance of Deep Neural Network inference becomes increasingly dependent on the result of quantization.", "Despite plenty of prior work on the quantization of weights or activations for neural networks, there is still a wide gap between the software quantizers and the low-precision accelerator implementation, which degrades either the efficiency of networks or that of the hardware for the lack of software and hardware coordination at design-phase.", "In this paper, we propose a learned linear symmetric quantizer for integer neural network processors, which not only quantizes neural parameters and activations to low-bit integer but also accelerates hardware inference by using batch normalization fusion and low-precision accumulators (e.g., 16-bit) and multipliers (e.g., 4-bit).", "We use a unified way to quantize weights and activations, and the results outperform many previous approaches for various networks such as AlexNet, ResNet, and lightweight models like MobileNet while keeping friendly to the accelerator architecture.", "Additional, we also apply the method to object detection models and witness high performance and accuracy in YOLO-v2.", "Finally, we deploy the quantized models on our specialized integer-arithmetic-only DNN accelerator to show the effectiveness of the proposed quantizer.", "We show that even with linear symmetric quantization, the results can be better than asymmetric or non-linear methods in 4-bit networks.", "In evaluation, the proposed quantizer induces less than 0.4\\% accuracy drop in ResNet18, ResNet34, and AlexNet when quantizing the whole network as required by the integer processors." ]

[ 1, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.3499999940395355, 0.2181818187236786, 0.10169491171836853, 0.12244897335767746, 0.05882352590560913, 0.17142856121063232, 0.10526315122842789, 0.04651162400841713 ]

[ "We introduce an efficient quantization process that allows for performance acceleration on specialized integer-only neural network accelerator." ]

[ "The prohibitive energy cost of running high-performance Convolutional Neural Networks (CNNs) has been limiting their deployment on resource-constrained platforms including mobile and wearable devices.", "We propose a CNN for energy-aware dynamic routing, called the EnergyNet, that achieves adaptive-complexity inference based on the inputs, leading to an overall reduction of run time energy cost without noticeably losing (or even improving) accuracy.", "That is achieved by proposing an energy loss that captures both computational and data movement costs.", "We combine it with the accuracy-oriented loss, and learn a dynamic routing policy for skipping certain layers in the networks, that optimizes the hybrid loss. ", "Our empirical results demonstrate that, compared to the baseline CNNs, EnergyNetcan trim down the energy cost up to 40% and 65%, during inference on the CIFAR10 and Tiny ImageNet testing sets, respectively, while maintaining the same testing accuracies. ", "It is further encouraging to observe that the energy awareness might serve as a training regularization and can even improve prediction accuracy: our models can achieve 0.7% higher top-1 testing accuracy than the baseline on CIFAR-10 when saving up to 27% energy, and 1.0% higher top-5 testing accuracy on Tiny ImageNet when saving up to 50% energy, respectively.", "While deep learning-powered Internet of Things (IoT) devices promise to dramatically revolutionize the way we live and work by enhancing our ability to recognize, analyze, and classify the world around us, this revolution has yet to be unleashed due to many fundamental challenges.", "Edge devices, such as smart phones, smart sensors, drones and robots, have limited energy and computation resources since they are battery-powered and have a small form factor.", "On the other hand, high-performance Convolutional Neural Networks (CNNs) come at a cost of prohibitive energy consumption BID0 .", "The CNNs with the highest accuracy have hundreds of layers and tens of millions of parameters.", "When deployed in practice, such networks drain the battery very quickly BID1 .", "Recently, there have been a number of methods proposed to reduce energy cost in CNNs, while not hampering their predictive power.", "Most of them aim to reduce the model size or the number of computations BID2 BID3 BID4 BID5 BID6 BID7 BID8 BID9 BID10 BID11 .", "However, BID1 shows that a smaller model size and fewer operations might not necessarily lead to a lower energy cost.", "BID1 uses energy cost to guide the pruning process, where the layer with the highest energy cost is pruned first.", "BID12 formulates the CNN training process as an optimization problem under a certain energy budget constraint.", "While both methods BID1 BID12 show promising results towards pursuing more energy-efficient CNN models, they do not incorporate energy costs into the training loss function to explicitly learn a more energy-efficient model.", "Furthermore, once their model structures are learned from training, it can only be fixed during the inference time, and there is no room for input-dependent adaptivity.", "This paper proposes a new CNN model that combines energy cost with a dynamic routing strategy to enable adaptive energy-efficient inference.", "Our proposed model, termed as EnergyNet, is a gated CNN architecture which employs conditional computing to route the input data through the network Figure 1 : EnergyNet Structure: each green circle G indicates an RNN gate and each blue square under G indicates one block of layers in the base model.", "To reduce the energy cost, the RNN gates generate routing strategies dynamically for different input images.", "By sharing the parameters between all RNN gates, they will have only 0.04% of the energy cost of the base CNN model, which is negligible.", "In this specific example, only the first and third blocks get executed.in an efficient path.", "Built on a base network (such as ResNet-34 or ResNet-50 BID13 ), EnergyNet uses an additional gating network BID10 to decide whether the current input should skip certain layers in the network or not.", "It optimizes a weighted combination of an accuracy loss and an energy loss which captures both the computational and memory data movement costs, under which EnergyNet is trained to find the optimal routing policy to reduce the energy cost of the model without degrading the prediction accuracy.", "Our empirical results demonstrate that, compared to the base network without gating nor dynamic inference, EnergyNet can trim down the energy cost up to 40% and 65%, during inference on the CIFAR10 and Tiny ImageNet testing sets, respectively, while maintaining almost the same testing accuracy.", "Interestingly enough, we find the energy-aware EnergyNet can even achieve win-win, by simultaneously improving the prediction accuracy and saving energy, potentially due to its equivalent effect as a training regularization to avoid overfitting.", "For example, our models achieve 0.7% higher top-1 testing accuracy than the baseline on CIFAR-10 when saving up to 27% energy, and 1.0% higher top-5 accuracy on Tiny ImageNet when saving up to 50% energy, respectively." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.09090908616781235, 0.290909081697464, 0.1111111044883728, 0.22727271914482117, 0.15094339847564697, 0.11940298229455948, 0.03448275476694107, 0.09302324801683426, 0.15789473056793213, 0.05882352590560913, 0, 0.19512194395065308, 0.0952380895614624, 0.307692289352417, 0.2222222238779068, 0.1666666567325592, 0.23999999463558197, 0.08695651590824127, 1, 0.12121211737394333, 0.11428570747375488, 0.1395348757505417, 0, 0.07999999821186066, 0.2181818187236786, 0.16949151456356049, 0.07843136787414551, 0.03999999538064003 ]

[ "This paper proposes a new CNN model that combines energy cost with a dynamic routing strategy to enable adaptive energy-efficient inference." ]

[ "Log-linear models models are widely used in machine learning, and in particular are ubiquitous in deep learning architectures in the form of the softmax.", "While exact inference and learning of these requires linear time, it can be done approximately in sub-linear time with strong concentrations guarantees.", "In this work, we present LSH Softmax, a method to perform sub-linear learning and inference of the softmax layer in the deep learning setting.", "Our method relies on the popular Locality-Sensitive Hashing to build a well-concentrated gradient estimator, using nearest neighbors and uniform samples.", "We also present an inference scheme in sub-linear time for LSH Softmax using the Gumbel distribution.", "On language modeling, we show that Recurrent Neural Networks trained with LSH Softmax perform on-par with computing the exact softmax while requiring sub-linear computations.", "Deep neural networks have achieved impressive successes in tasks spanning vision BID9 BID16 , language BID3 , speech BID6 BID27 and videos BID1 .", "While these models can vastly differ in architecture, activation functions, and presence of recurrence, they (almost) all share a common trait: the softmax layer.", "The softmax layer, or log-linear model, is a widely used model in machine learning and statistics that transforms a feature vector into a distribution over the output space, modeling log-probabilities as a linear function of the feature vector.", "For example, in object classification, the softmax layer at the end of a deep convolutional network transforms a feature vector into a probability distribution over classes for the image; in language modeling using recurrent neural networks, it maps the hidden state to a distribution over next words.While parameterizing for logits offers modeling flexibility, inference and learning have linear runtime in the number of classes.", "Indeed, both of these require computing the un-normalized probability for every class to compute the partition function and retrieve an actual probability distribution.", "Problems with large output spaces arise naturally in many areas like natural language processing (NLP), where the output space is a language's vocabulary and can be on the order of hundreds of thousands of elements BID15 ; BID12 .", "This can also occur in computer vision BID14 when attempting tag prediction on massive, weakly-labeled datasets such as Flickr100M BID31 .Many", "solutions have been proposed to address this bottleneck, all revolving around two themes: approximation of the softmax probabilities or computation of exact probabilities for an approximate model. Canonical", "examples of the former are importance sampling (IS) or noise contrastive estimation (NCE; BID8 ). Instead of", "computing probabilities over the whole output space, these methods compute the softmax over a smaller, sampled vocabulary and re-weight the probabilities, providing an unbiased estimator. An illustration", "of the latter is Hierarchical Softmax BID24 , where the output classes are first clustered such that you only need to compute the softmax over a smaller output space. While the former", "is an unbiased estimate, it comes with no concentration guarantees, and it is often more art than science to craft proposal distributions which will provide low-variance estimators. The latter, while", "efficient, requires carefully hand-crafted clustering of the output space, at the risk of making mistakes from which there is no recovery.More recently, estimators based on nearest neighbor search have been proposed for inference and learning in log-linear models BID25 BID26 . These estimators", "hinge on Maximum Inner Product Search using Locality-Sensitive to retrieve the largest logits of the distribution and account for the tail with uniformly sampled classes. They boast strong", "theoretical guarantees and well-established concentration bounds. However, they were", "constrained to toy settings and not directly applicable to real-world, large-scale, machine learning. In this work, we build", "upon these estimators to make them amenable to deep learning practitioners, without losing any theoretical guarantees. We first show how they", "can be extended to be usable within training of deep learning models, then present our efficient implementation, adapted to deep learning hardware and frameworks. Finally, we show the applicability", "and efficiency of our method by evaluating on a real-world task: language modeling. We show significant perplexity gains", "against competing methods with significant speed-ups.Our contributions are as follows:• We present a new deep learning layer, LSH Softmax, an efficient replacement for the softmax layer based on Locality-Sensitive Hashing and the Gumbel distribution, for any deep learning architecture, with strong theoretical guarantees for sub-linear learning and inference.• We provide details for efficient implementation", "on deep learning hardware (GPUs) and modern deep learning frameworks BID0 BID19 ).• Empirically, we show, on several datasets, that", "training and sampling from LSH Softmax performs similarly to an exact softmax while requiring significantly less FLOPS.", "In this work, we presented LSH Softmax, a softmax approximation layer for large output spaces with sub-linear learning and inference cost (in the number of states) and strong theoretical guarantees.", "We showcased both its applicability and efficiency by evaluating LSH on a common NLP task, language modeling.", "On several datasets for this task, we report perplexity closest to exact training among all baselines, as well as significant speed-ups.", "Our hope is that, for any architecture, this layer could be chosen in lieu of softmax, when the output space is sufficiently large to warrant the approximation.To that end, we plan to release source-code with the camera-ready version." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0 ]

[ 0.12765957415103912, 0.23529411852359772, 0.4313725531101227, 0.12244897335767746, 0.2222222238779068, 0.23076923191547394, 0.07843136787414551, 0.15094339847564697, 0.16393442451953888, 0.2278480976819992, 0.11999999731779099, 0.158730149269104, 0.03999999538064003, 0.1090909019112587, 0, 0.11320754140615463, 0.07017543166875839, 0.07017543166875839, 0.14492753148078918, 0.1818181723356247, 0.10526315122842789, 0.1304347813129425, 0.08163265138864517, 0.18518517911434174, 0.42553192377090454, 0.4054054021835327, 0.1702127605676651, 0.1304347813129425, 0.517241358757019, 0.52173912525177, 0.08163265138864517, 0.15625 ]

[ "we present LSH Softmax, a softmax approximation layer for sub-linear learning and inference with strong theoretical guarantees; we showcase both its applicability and efficiency by evaluating on a real-world task: language modeling." ]

[ "Can the success of reinforcement learning methods for simple combinatorial optimization problems be extended to multi-robot sequential assignment planning?", "In addition to the challenge of achieving near-optimal performance in large problems, transferability to an unseen number of robots and tasks is another key challenge for real-world applications.", "In this paper, we suggest a method that achieves the first success in both challenges for robot/machine scheduling problems.\n \n", "Our method comprises of three components.", "First, we show any robot scheduling problem can be expressed as a random probabilistic graphical model (PGM).", "We develop a mean-field inference method for random PGM and use it for Q-function inference.", "Second, we show that transferability can be achieved by carefully designing two-step sequential encoding of problem state.", "Third, we resolve the computational scalability issue of fitted Q-iteration by suggesting a heuristic auction-based Q-iteration fitting method enabled by transferability we achieved.\n \n", "We apply our method to discrete-time, discrete space problems (Multi-Robot Reward Collection (MRRC)) and scalably achieve 97% optimality with transferability.", "This optimality is maintained under stochastic contexts.", "By extending our method to continuous time, continuous space formulation, we claim to be the first learning-based method with scalable performance in any type of multi-machine scheduling problems; our method scalability achieves comparable performance to popular metaheuristics in Identical parallel machine scheduling (IPMS) problems.", "Suppose that we are given a set of robots and seek to serve a set of spatially distributed tasks.", "A reward is given for serving each task promptly -resulting in a time-decaying reward collection problem -or when completing the entire set of tasks -resulting in a makespan minimization problem.", "As the capability to control and route individual robots has increased [Li (2017) ], efficient orchestration of robots arises as an important remaining concern for such problems.", "Multi-robot planning problems.", "In this paper, we focus on orchestration problems that can be formulated as robot planning problems.", "A key assumption in such orchestration problems is that we are given information on the \"duration of time required for an assigned robot to complete a task\".", "This duration may be deterministic (e.g. as in a Traveling Salesman Problem (TSP) or Vehicle Routing Problem (VRP)) or random with given probability distribution (c.f., [Omidshafiei et al. (2017) ]).", "1 .", "We call this duration the task completion time.", "Due to their combinatorial nature, robot planning problems suffer from exponential computational complexity.", "Even in the context of single-robot scheduling problems (e.g., TSP) scalability is a concern.", "Planning for multiple robots exacerbates the scalability issue.", "While scalable heuristic methods have been developed for various deterministic multi-robot planning problems (c.f., [Rossi Proposed methods. In the seminal paper [Dai et al. (2017) ], the authors observed that combinatorial optimization problems such as TSP can be formulated as sequential decision making problems.", "Decision making in such a sequential framework relies on an estimate of future costs Q(s, a) for an existing task sequence s and candidate next task a.", "With this estimate, given the prior decisions s at each decision step, they select the next task a to minimize the future cost estimate.", "[Dai et al. (2017) ]'s solution framework relies on the following three assumptions.", "1) For each combinatorial optimization problem, one can heuristically choose how to induce a graph representation of (s, a).", "In the case of TSP, the paper induces a fully connected graph for every possible next task.", "2) This induced graph representation can be considered as a probabilistic graphical model (PGM) [Koller & Friedman (2009) ].", "This PGM can be used with a graph-based mean-field inference method called structure2vec [Dai et al. (2016) ] to infer Q(s, a) for use in combinatorial optimization problems.", "3) Inference of Q(s, a) can be learned by the reinforcement framework called fitted Q-iteration.", "We create a solution framework to achieve scalability and transferability for multi-robot planning that builds in numerous directions upon the foundation of [Dai et al. (2017) ] as follows: 1.", "State representation and mean-field inference theory for random PGM.", "Instead of heuristically inducing a PGM, we show that a robot scheduling problem exactly induces a random PGM.", "Since there exists no mean-field inference theory for random PGM, we develop the theory and corresponding new structure2vec iteration.", "2. Sequential encoding of information for transferability.", "To achieve transferability in terms of the number of robots and tasks, we carefully design a two-step hierarchical mean-field inference [Ranganath et al. (2015) ].", "Each step is designed to infer certain information.", "The first step is designed to infer each task's relative graphical distance from the robots.", "The second step is designed to infer Q(s, a) (a here refers to a joint assignment of robots).", "While the first step is by its nature transferable to any number of tasks and robots, the transferability in inference of the second step is achieved by the scale-free characteristic of fitted Q-iteration [van Hasselt et al. (2015) ].", "That is, the relative magnitudes of Q(s, a) values are sufficient to select an action a.", "3. Auction-based assignment.", "Even if we can infer Q(s, a) precisely, the computation time required to select an action a using the maximum Q(s, a) operation exponentially increases as robots and tasks increase.", "To resolve this issue, we suggest a heuristic auction that is enabled by the transferability of our Q(s, a) inference.", "Even though this heuristic auction selects a with only polynomial computational complexity, it provides surprisingly good choices for a.", "(In fact, this heuristic auction increases the performance empirically relative to using the max operation.) time τ i to complete -we call this the processsing time.", "This time is the same independent of which machine serves the task.", "We incorporate one popular extension and allow 'sequence-dependent setup times'.", "In this case, a machine must conduct a setup prior to serving each task.", "The duration of this setup depends on the current task i and the task j that was previously served on that machine -we call this the setup time.", "The completion time for each task is thus the sum of the setup time and processing time.", "Under this setting, we solve the IPMS problem for make-span minimization as discussed in [Kurz et al. (2001) ].", "That is, we seek to minimize the total time spent from the start time to the completion of the last task.", "The IPMS formulation resembles our MRRC formulation in continuous-time and continuous-space and we relegate the detailed formulation to Appendix B.", "We presented a learning-based method that achieves the first success for multi-robot/machine scheduling problems in both challenges: scalable performance and tranferability.", "We identified that robot scheduling problems have an exact representation as random PGM.", "We developed a meanfield inference theory for random PGM and extended structure2vec method of Dai et al. (2016) .", "To overcome the limitations of fitted Q-iteration, a heuristic auction that was enabled by transferability is suggested.", "Through experimental evaluation, we demonstrate our method's success for MRRC problems under a deterministic/stochastic environment.", "Our method also claims to be the first learning-based algorithm that achieves scalable performance among machine scheduling algorithms; our method achieves a comparable performance in a scalable manner.", "Our method for MRRC problems can be easily extended to ride-sharing problems or package delivery problems.", "Given a set of all user requests to serve, those problems can be formulated as a MRRC problem.", "For both ride-sharing and package delivery, it is reasonable to assume that the utility of a user depends on when she is completely serviced.", "We can model how the utility of a user decreases over time since when it appears and set the objective function of problems as maximizing total collected user utility.", "Now consider a task 'deliver user (or package) from A to B'.", "This is actually a task \"Move to location A and then move to location B\".", "If we know the completion time distribution of each move (as we did for MRRC), the task completion time is simply the sum of two random variables corresponding to task completion time distribution of the moves in the task.", "Indeed, ride-sharing or package delivery problems are of such tasks (We can ignore charging moves for simplicity, and also we don't have to consider simple relocation of vehicles or robots since we don't consider random customer arrivals).", "Therefore, both ride-sharing problems and package delivery problems can be formulated as MRRC problems.", "A MRRC WITH CONTINUOUS STATE/CONTINUOUS TIME SPACE FORMULATION, OR WITH SETUP TIME AND PROCESSING TIME", "In continuous state/continuous time space formulation, the initial location and ending location of robots and tasks are arbitrary on R 2 .", "At every moment at least a robot finishes a task, we make assignment decision for a free robot(s).", "We call this moments as 'decision epochs' and express them as an ordered set (t 1 , t 2 , . . . , t k , . . . ).", "Abusing this notation slightly, we use (·) t k = (·) k .", "Task completion time can consist of three components: travel time, setup time and processing time.", "While a robot in the travel phase or setup phase may be reassigned to other tasks, we can't reassign a robot in the processing phase.", "Under these assumptions, at each decision epoch robot r i is given a set of tasks it can assign itself: if it is in the traveling phase or setup phase, it can be assigned to any tasks or not assigned; if it is in the processing phase, it must be reassigned to its unfinished task.", "This problem can be cast as a Markov Decision Problem (MDP) whose state, action, and reward are defined as follows:", "R k is the set of all robots and T k is the set of all tasks; The set of directed edges", "where a directed edge", "is a random variable which denotes task completion time of robot i in R k to service task j in T k and a directed edge titj ∈ E T T k denotes a task completion time of a robot which just finished serving task i in T k to service task j in T k .", "E RT k contains information about each robot's possible assignments:", ", where E ri t is a singleton set if robot i is in the processing phase and it must be assigned to its unfinished task, and otherwise it is the set of possible assignments from robot r i to remaining tasks that are not in the processing phase.", "Action.", "The action a k at decision epoch k is the joint assignment of robots given the current state s k = G k .", "The feasible action should satisfy the two constraints: No two robots can be assigned to a task; some robots may not be assigned when number of robots are more than remaining tasks.", "To best address those restrictions, we define an action a k at time t as a maximal bipartite matching in bipartite sub-graph ((R k ∪ T k ), E RT k ) of graph G k .", "For example, robot i in R k is matched with task j in T k in an action a k if we assign robot i to task j at decision epoch t.", "We denote the set of all possible actions at epoch k as A k .", "Reward.", "In MRRC, Each task has an arbitrarily determined initial age.", "At each decision epoch, the age of each task increases by one.", "When a task is serviced, a reward is determined only by its age when serviced.", "Denote this reward rule as R(k).", "One can easily see that whether a task is served at epoch k is completely determined by s k , a k and s k+1 .", "Therefore, we can denote the reward we get with s k , a k and s k+1 as R(s k , a k , s k+1 ).", "Objective.", "We can now define an assignment policy φ as a function that maps a state s k to action a k .", "Given s 0 initial state, an MRRC problem can be expressed as a problem of finding an optimal assignment policy φ * such that" ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.06451612710952759, 0.05405404791235924, 0.0624999962747097, 0, 0.06896550953388214, 0.07999999821186066, 0.06896550953388214, 0, 0.1875, 0, 0.04255318641662598, 0.0714285671710968, 0, 0.10526315122842789, 0.13333332538604736, 0.14814814925193787, 0.05128204822540283, 0, 0, 0.07999999821186066, 0.0714285671710968, 0, 0.07547169178724289, 0.0555555522441864, 0, 0, 0.12903225421905518, 0.0714285671710968, 0.12903225421905518, 0.09999999403953552, 0.07407406717538834, 0.0476190447807312, 0.0952380895614624, 0, 0.06666666269302368, 0, 0.0555555522441864, 0, 0, 0, 0.04651162400841713, 0, 0, 0.1538461446762085, 0, 0, 0.05882352590560913, 0, 0.09090908616781235, 0, 0.060606054961681366, 0.07692307233810425, 0.06451612710952759, 0, 0.06896550953388214, 0.12121211737394333, 0.07999999821186066, 0.06666666269302368, 0, 0.07407406717538834, 0, 0.1538461446762085, 0.13793103396892548, 0.05714285373687744, 0.1621621549129486, 0, 0.07999999821186066, 0, 0.13636362552642822, 0.25, 0, 0.06451612710952759, 0, 0.060606054961681366, 0, 0.1599999964237213, 0, 0.038461536169052124, 0.12903225421905518, 0.07692307233810425, 0, 0.04999999701976776, 0, 0.0416666641831398, 0, 0.05128204822540283, 0.0476190447807312, 0, 0, 0, 0, 0, 0, 0.1249999925494194, 0.13793103396892548, 0.06666666269302368, 0.05882352590560913 ]

[ "RL can solve (stochastic) multi-robot/scheduling problems scalably and transferably using graph embedding" ]

[ "Curriculum learning consists in learning a difficult task by first training on an easy version of it, then on more and more difficult versions and finally on the difficult task.", "To make this learning efficient, given a curriculum and the current learning state of an agent, we need to find what are the good next tasks to train the agent on.\n", "Teacher-Student algorithms assume that the good next tasks are the ones on which the agent is making the fastest progress or digress.", "We first simplify and improve them.", "However, two problematic situations where the agent is mainly trained on tasks it can't learn yet or it already learnt may occur.\n", "Therefore, we introduce a new algorithm using min max ordered curriculums that assumes that the good next tasks are the ones that are learnable but not learnt yet.", "It outperforms Teacher-Student algorithms on small curriculums and significantly outperforms them on sophisticated ones with numerous tasks.", "Curriculum learning.", "An agent with no prior knowledge can learn a lot of tasks by reinforcement, i.e. by reinforcing (taking more often) actions that lead to higher reward.", "But, for some very hard tasks, it is impossible.", "Let's consider the following task:Figure 1: The agent (in red) receives a reward of 1 when it picks up the blue ball in the adjacent room.", "To do so, it has to first open the gray box, take the key inside and then open the locked door.", "This is an easy task for humans because we have prior knowledge: we know that a key can be picked up, that we can open a locked door with a key, etc...", "However, most of the time, the agent starts with no prior knowledge, i.e. it starts by acting randomly. Therefore, it has a probability near 0 of achieving the task in a decent number of time-steps, so it has a probability near 0 of getting reward, so it can't learn the task by reinforcement.One solution to still learn this task is to do curriculum learning BID0 ), i.e. to first train the agent on an easy version of the task, where it can get reward and learn, then train on more and more difficult versions using the previously learnt policy and finally, train on the difficult task.Learning by curriculum may be decomposed into two parts:1. Defining the curriculum, i.e. the set of tasks the agent may be trained on. 2. Defining the program, i.e. the sequence of curriculum's tasks it will be trained on.These two parts can be done online, during training.Curriculum learning algorithms. Defining a curriculum and a program can be done manually, e.g. by defining a hand-chosen performance threshold for advancement to the next task BID6 ; BID5 ).However, if an efficient algorithm is found, it may save us a huge amount of time in the future. Besides, efficient (and more efficient than humans) algorithms are likely to exist because they can easily mix in different tasks (what is hard for humans) and then:• avoid catastrophic forgetting by continuously retraining on easier tasks;• quickly detect learnable but not learnt yet tasks.Hence, it motivates the research of curriculum learning algorithms.Curriculum learning algorithms can be grouped into two categories:1. curriculum algorithms: algorithms that define the curriculum; 2. program algorithms: algorithms that define the program, i.e. that decide, given a curriculum and the learning state of the agent, what are the good next tasks to train the agent on.In this paper, we will focus on program algorithms, in the reinforcement learning context. Recently, several such algorithms emerged, focused on the notion of learning progress BID4 ; BID3 BID2 ). BID4 proposed four algorithms (called Teacher-Student) based on the assumption that the good next tasks are the ones on which the agent is making the fastest progress or digress.We first simplify and improve Teacher-Student algorithms (section 4). However, even improved, two problematic situations where the agent is mainly trained on tasks it can't learn or it already learnt may occur. Therefore, we introduce a new algorithm (section 5), focused on the notion of mastering rate, based on the assumption that the good next tasks are the ones that are learnable but not learnt yet.We show that this algorithm outperforms Teacher-Student algorithms on small curriculums and significantly outperforms them on sophisticated ones with numerous tasks." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.2857142686843872, 0.25, 0.20512819290161133, 0.07692307233810425, 0.0952380895614624, 0.22727271914482117, 0.17142856121063232, 0.17391303181648254, 0.06896550953388214, 0.13636362552642822, 0.052631575614213943, 0.1304347813129425, 0.14529913663864136 ]

[ "We present a new algorithm for learning by curriculum based on the notion of mastering rate that outperforms previous algorithms." ]

[ "The fields of artificial intelligence and neuroscience have a long history of fertile bi-directional interactions.", "On the one hand, important inspiration for the development of artificial intelligence systems has come from the study of natural systems of intelligence, the mammalian neocortex in particular.", "On the other, important inspiration for models and theories of the brain have emerged from artificial intelligence research.", "A central question at the intersection of these two areas is concerned with the processes by which neocortex learns, and the extent to which they are analogous to the back-propagation training algorithm of deep networks.", "Matching the data efficiency, transfer and generalisation properties of neocortical learning remains an area of active research in the field of deep learning.", "Recent advances in our understanding of neuronal, synaptic and dendritic physiology of the neocortex suggest new approaches for unsupervised representation learning, perhaps through a new class of objective functions, which could act alongside or in lieu of back-propagation.", "Such local learning rules have implicit rather than explicit objectives with respect to the training data, facilitating domain adaptation and generalisation. ", "Incorporating them into deep networks for representation learning could better leverage unlabelled datasets to offer significant improvements in data efficiency of downstream supervised readout learning, and reduce susceptibility to adversarial perturbations, at the cost of a more restricted domain of applicability.\n" ]

[ 0, 0, 0, 0, 1, 0, 0, 0 ]

[ 0.07407406717538834, 0.11428570747375488, 0.13333332538604736, 0.1428571343421936, 0.1875, 0.1304347813129425, 0.17142856121063232, 0.11538460850715637 ]

[ "Inspiration from local dendritic processes of neocortical learning to make unsupervised learning great again." ]

[ "Real-world Question Answering (QA) tasks consist of thousands of words that often represent many facts and entities.", "Existing models based on LSTMs require a large number of parameters to support external memory and do not generalize well for long sequence inputs.", "Memory networks attempt to address these limitations by storing information to an external memory module but must examine all inputs in the memory.", "Hence, for longer sequence inputs the intermediate memory components proportionally scale in size resulting in poor inference times and high computation costs.\n\n", "In this paper, we present Adaptive Memory Networks (AMN) that process input question pairs to dynamically construct a network architecture optimized for lower inference times.", "During inference, AMN parses input text into entities within different memory slots.", "However, distinct from previous approaches, AMN is a dynamic network architecture that creates variable numbers of memory banks weighted by question relevance.", "Thus, the decoder can select a variable number of memory banks to construct an answer using fewer banks, creating a runtime trade-off between accuracy and speed. \n\n", "AMN is enabled by first, a novel bank controller that makes discrete decisions with high accuracy and second, the capabilities of a dynamic framework (such as PyTorch) that allow for dynamic network sizing and efficient variable mini-batching.", "In our results, we demonstrate that our model learns to construct a varying number of memory banks based on task complexity and achieves faster inference times for standard bAbI tasks, and modified bAbI tasks.", "We achieve state of the art accuracy over these tasks with an average 48% lower entities are examined during inference.", "Question Answering (QA) tasks are gaining significance due to their widespread applicability to recent commercial applications such as chatbots, voice assistants and even medical diagnosis BID7 ).", "Furthermore, many existing natural language tasks can also be re-phrased as QA tasks.", "Providing faster inference times for QA tasks is crucial.", "Consumer device based question-answer services have hard timeouts for answering questions.", "For example, Amazon Alexa, a popular QA voice assistant, allows developers to extend the QA capabilities by adding new \"Skills\" as remote services BID0 ).", "However, these service APIs are wrapped around hard-timeouts of 8 seconds which includes the time to transliterate the question to text on Amazon's servers and the round-trip transfer time of question and the answer from the remote service, and sending the response back to the device.", "Furthermore, developers are encouraged to provide a list of questions (\"utterances\") apriori at each processing step to assist QA processing BID0 ).Modeling", "QA tasks with LSTMs can be computationally expensive which is undesirable especially during inference. Memory networks", ", a class of deep networks with explicit addressable memory, have recently been used to achieve state of the art results on many QA tasks. Unlike LSTMs,", "where the number of parameters grows exponentially with the size of memory, memory networks are comparably parameter efficient and can learn over longer input sequences. However, they", "often require accessing all intermediate memory to answer a question. Furthermore,", "using focus of attention over the intermediate state using a list of questions does not address this problem. Soft attention", "based models compute a softmax over all states and hard attention models are not differentiable and can be difficult to train over a large state space. Previous work", "on improving inference over memory networks has focused on using unsupervised clustering methods to reduce the search space BID2 ; BID19 ). Here, the memory", "importance is not learned and the performance of nearest-neighbor style algorithms is often comparable to a softmax operation over memories. To provide faster", "inference for long sequence-based inputs, we present Adaptive Memory Networks (AMN), that constructs a memory network on-the-fly based on the input. Like past approaches", "to addressing external memory, AMN constructs the memory nodes dynamically. However, distinct from", "past approaches, AMN constructs a memory architecture with network properties that are decided dynamically based on the input story. Given a list of possible", "questions, our model computes and stores the entities from the input story in a memory bank. The entities represent the", "hidden state of each word in the story while a memory bank is a collection of entities that are similar w.r.t the question. As the number of entities", "grow, our network learns to construct new memory banks and copies entities that are more relevant towards a single bank. Entities may reside in different", "bank depending on their distance from the question. Hence, by limiting the decoding", "step to a dynamic number of constructed memory banks, AMN achieves lower inference times. AMN is an end-to-end trained model", "with dynamic learned parameters for memory bank creation and movement of entities.Figure 1 demonstrates a simple QA task where AMN constructs two memory banks based on the input. During inference only the entities", "in the left bank are considered reducing inference times. To realize its goals, AMN introduces", "a novel bank controller that uses reparameterization trick to make discrete decisions with high accuracy while maintaining differentiability. Finally, AMN also models sentence structures", "on-the-fly and propagates update information for all entities that allows it to solve all 20 bAbI tasks.", "In this paper, we present Adaptive Memory Network that learns to adaptively organize the memory to answer questions with lower inference times.", "Unlike NTMs which learn to read and write at individual memory locations, Adaptive Memory Network demonstrates a novel design where the learned memory management is coarse-grained that is easier to train.Through our experiments, we demonstrate that AMN can learn to reason, construct, and sort memory banks based on relevance over the question set.AMN architecture is generic and can be extended to other types of tasks where the input sequence can be separated into different entities.", "In the future, we plan to evaluate AMN over such tasks to evaluate AMN generality.", "We also plan to experiment with larger scale datasets (beyond bAbI, such as a document with question pairs) that have a large number of entities to further explore scalability.Method Complexity Entnet BID12 We describe our overall algorithm in pseudo-code in this section.", "We follow the notation as described in the paper.", "DISPLAYFORM0 Algorithm 1 AMN(S, q, a) DISPLAYFORM1 for word w ∈ s do 4: DISPLAYFORM2 end for 6: DISPLAYFORM3 for memory bank m i ∈ M do 8: DISPLAYFORM4 n mi ← SGRU(D, n mi ) We compare the computations costs during the decode operation during inference for solving the extended bAbi task.", "We compute the overheads for AMN Entnet BID12 ) and GGT-NN.", "TAB2 gives the decode comparisons between AMN, Entnet and GGT-NN.", "Here, |V | represents to the total number of entities for all networks.", "GGT-NN can dynamically create nodes and k k is hyper parameter the new nodes created for S sentences in input story.", "α is the percent of entities stored in the final bank w.r.t to the total entities for AMN.We compare the wall clock execution times for three tasks within bAbI for 1000 examples/task.", "We compare the wall-clock times for three tasks.", "We compare the inference times of considering all banks (and entities) versus the just looking at the passing banks as required by AMN.", "We find that AMN requires fewer banks and as a consequence fewer entities and saves inference times.", "In this section, we understand memory bank behavior of AMN.", "Figure 3 shows the memory banks and the entity creation for a single story example, for some of the tasks from bAbI.", "Depending upon the task, and distance from the question AMN creates variable number of memory banks.", "The heatmap demonstrates how entities are copied across memory banks.", "Grey blocks indicate absence of those banks.Under review as a conference paper at ICLR 2018 Figure 4 shows how propagation happens after every time step.", "The nodes represent entities corresponding to words in a sentence.", "As sentences are processed word by word, a directed graph is drawn progressively from w 0 ...w i ...w N .", "If sentence l k 's path contains nodes already in the current directed graph, l k will include said nodes in the its path. After l k is added to A, the model propagates the new update hidden state information a i among all node states using a GRU. a i for each node i is equal to the sum of the incoming edges' node hidden states.", "Additionally, we add a particular emphasis on l k to simulate recency.", "At face value, one propagation step of A will only have a reachability of its immediate neighbor, so to reach all nodes, A is raised to a consecutive power r to reach and update each intermediate node.", "r can be either the longest path in A or a set parameter." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0, 0, 0.1538461446762085, 0.07407407462596893, 0.13333332538604736, 0, 0, 0, 0.052631575614213943, 0.1111111119389534, 0.1599999964237213, 0, 0, 0.2857142686843872, 0, 0, 0, 0, 0.380952388048172, 0.12903225421905518, 0.13333332538604736, 0, 0, 0, 0.14814814925193787, 0.07407407462596893, 0.13793103396892548, 0, 0.0714285671710968, 0, 0, 0, 0, 0.0833333283662796, 0.1111111119389534, 0.09999999403953552, 0.06896551698446274, 0, 0.23076923191547394, 0.03124999813735485, 0, 0.04651162400841713, 0, 0.0416666641831398, 0, 0, 0.1111111044883728, 0, 0, 0, 0.07999999821186066, 0.09999999403953552, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ "Memory networks with faster inference" ]

[ "When a bilingual student learns to solve word problems in math, we expect the student to be able to solve these problem in both languages the student is fluent in, even if the math lessons were only taught in one language.", "However, current representations in machine learning are language dependent.", "In this work, we present a method to decouple the language from the problem by learning language agnostic representations and therefore allowing training a model in one language and applying to a different one in a zero shot fashion.", "We learn these representations by taking inspiration from linguistics, specifically the Universal Grammar hypothesis and learn universal latent representations that are language agnostic (Chomsky, 2014; Montague, 1970).", "We demonstrate the capabilities of these representations by showing that the models trained on a single language using language agnostic representations achieve very similar accuracies in other languages.", "Anecdotally speaking, fluent bilingual speakers rarely face trouble translating a task learned in one language to another.", "For example, a bilingual speaker who is taught a math problem in English will trivially generalize to other known languages.", "Furthermore there is a large collection of evidence in linguistics arguing that although separate lexicons exist in multilingual speakers the core representations of concepts and theories are shared in memory BID1 BID28 BID6 .", "The fundamental question we're interested in answering is on the learnability of these shared representations within a statistical framework.We approached this problem from a linguistics perspective.", "Languages have vastly varying syntactic features and rules.", "Linguistic Relativity studies the impact of these syntactic variations on the formations of concepts and theories BID5 .", "Within this framework of study, the two schools of thoughts are linguistic determinism and weak linguistic influence.", "Linguistic determinism argues that language entirely forms the range of cognitive processes, including the creation of various concepts, but is generally agreed to be false BID18 BID5 .", "Although there exists some weak linguistic influence, it is by no means fundamental BID0 .", "The superfluous nature of syntactic variations across languages brings forward the argument of principles and parameters (PnP) which hypothesizes the existence of a small distributed parameter representation that captures the syntactic variance between languages denoted by parameters (e.g. head-first or head-final syntax), as well as common principles shared across all languages BID12 .", "Universal Grammar (UG) is the study of principles and the parameters that are universal across languages BID29 .The", "ability to learn these universalities would allow us to learn representations of language that are fundamentally agnostic of the specific language itself. Doing", "so would allow us to learn a task in one language and reap the benefits of all other languages without needing multilingual datasets. We take", "inspiration from the UG hypothesis and learn latent representations that are language agnostic which allow us to solve downstream problems in new languages without the need of any language-specific training data. We do not", "make any claims about the Universal Grammar hypothesis, but simply take inspiration from it.", "Universal Grammar also comments on the learnability of grammar, stating that statistical information alone is not enough to learn grammar and some form of native language faculty must exist, sometimes titled the poverty of stimulus (POS) argument BID10 BID23 ).", "The goal of our paper is not to make a statement on the Universal Grammar hypothesis.", "But from a machine learning perspective, we're interested in extracting informative features.", "That being said it is of interest to what extent language models capture grammar and furthermore the extent to which models trained with our objective learn grammar.", "One way to measure universality is by studying perplexity of our multi-lingual language model as we increase the number of languages.", "To do so we trained 6 UG-WGAN models on the following languages: English, Russian, Arabic, Chinese, German, Spanish, French.", "We maintain the same procedure as described above.", "The hidden size of the language model was increased to 1024 with 16K BPE tokens being used.", "The first model was trained on English Russian, second was trained on English Russian Arabic and so on.", "For Arabic we still trained from left to right even though naturally the language is read from right to left.", "We report the results in FIG3 .", "As we increase the number of languages the perplexity gap between constrained and unconstrained UG-WGAN (λ = 0.0) decreases which implies while controlling capacity, our constrained (universal λ = 0.1) language model, models language (almost) as well as jointly trained language models with no universal constraints (λ = 0.0).Furthermore", ", the heatmap in FIG3 shows the perplexity gap of UG-WGAN trained on any combination of 2 languages from our set of 7. We can treat", "the perplexities as a loose measure of distance λ = 0.0 λ = 0.1 en earth's oxide is a monopoly that occurs towing of the carbon-booed trunks, resulting in a beam containing of oxygen through the soil, salt, warm waters, and the different proteins.the practice of epimatic behaviours may be required in many ways of all non-traditional entities.the groove and the products are numeric because they are called \"pressibility\" (ms) nutrients containing specific different principles that are available from the root of their family, including a wide variety of molecular and biochemical elements. a state line", "is a self-government environment for statistical cooperation, which is affected by the monks of canada, the east midland of the united kingdom.however, compared to the listing of special definitions, it has evolved to be congruent with structural introductions, allowing to form the chemical form.the vernacular concept of physical law is not as an objection (the whis) but as a universal school.es la revista ms reciente vari el manuscrito originalmente por primera vez en la revista publicada en 1994.en el municipio real se localiza al mar del norte y su entorno en escajros alto, con mayor variedad de cclica poblacin en forma de cerca de 1070 km2.de hecho la primera cancin de \"blebe cantas\", pahka zanjiwtryinvined cot de entre clases de fanticas, apareci en el ornitlogo sello triusion, jr., en la famosa publicacin playboy de john allen.fue el ltimo habitantes de suecia, con tres hijos, atasaurus y aminkinano (nuestra).The names of", "large predators in charlesosaurus include bird turtles hibernated by aerial fighters and ignored fish.jaime en veracruz fue llamado papa del conde mayor de valdechio, hijo de diego de ziga. We see from", "Figure 2 that perplexity worsens proportional to λ. We explore", "the differences by sampling sentences from an unconstrained language model and λ = 0.1 language model trained towards English and Spanish in Table 3 . In general", "there is a very small difference between a language model trained with our objective and one without. The constrained", "model tends to make more gender mistakes and mistakes due to Plural-Singular Form in Spanish. In English we saw", "virtually no fundamental differences between the language models. One explanation of", "this phenomena comes from the autonomy of syntax argument, which argues that semantics have no weight on syntax BID16 . Our hypothesis is", "that both models learn syntax well, but the models with better perplexity generate sentences with better or clearer semantic meaning. Although completely", "learning grammar from statistical signals might be improbable, we can still extract useful information.", "In this paper we introduced an unsupervised approach toward learning language agnostic universal representations by taking inspiration from the Universal Grammar hypothesis.", "We showed that we can use these representations to learn tasks in one language and automatically transfer them to others with no additional training.", "Furthermore we studied the importance of the Wasserstein constraint through the λ hyper-parameter.", "And lastly we explored the difference between a standard multi-lingual language model and UG-WGAN by studying the generated outputs of the respective language models as well as the perplexity gap growth with respect to the number of languages." ]

[ 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.17543859779834747, 0.1764705777168274, 0.29629629850387573, 0.5199999809265137, 0.19999998807907104, 0.0952380895614624, 0.09090908616781235, 0.072727270424366, 0.11764705181121826, 0, 0.04999999701976776, 0.04999999701976776, 0.11999999731779099, 0, 0.11940298229455948, 0.2857142686843872, 0.27272728085517883, 0.19999998807907104, 0.37931033968925476, 0.307692289352417, 0.19354838132858276, 0.19512194395065308, 0.10810810327529907, 0.2083333283662796, 0.2222222238779068, 0.13636362552642822, 0.060606058686971664, 0.1428571343421936, 0, 0.24390242993831635, 0.06451612710952759, 0.1818181723356247, 0.1666666567325592, 0.0416666641831398, 0.0559440515935421, 0.0357142798602581, 0.05714285373687744, 0.12244897335767746, 0.04651162400841713, 0.0952380895614624, 0.1111111044883728, 0.1304347813129425, 0.09090908616781235, 0.20512820780277252, 0.5106382966041565, 0.25, 0.1111111044883728, 0.178571417927742 ]

[ "By taking inspiration from linguistics, specifically the Universal Grammar hypothesis, we learn language agnostic universal representations which we can utilize to do zero-shot learning across languages." ]

[ "Generative models with both discrete and continuous latent variables are highly motivated by the structure of many real-world data sets.", "They present, however, subtleties in training often manifesting in the discrete latent variable not being leveraged.", "In this paper, we show why such models struggle to train using traditional log-likelihood maximization, and that they are amenable to training using the Optimal Transport framework of Wasserstein Autoencoders.", "We find our discrete latent variable to be fully leveraged by the model when trained, without any modifications to the objective function or significant fine tuning.", "Our model generates comparable samples to other approaches while using relatively simple neural networks, since the discrete latent variable carries much of the descriptive burden.", "Furthermore, the discrete latent provides significant control over generation.", "Unsupervised learning using generative latent variable models provides a powerful and general approach to learning the underlying, low-dimensional structure from large, unlabeled datasets.", "Perhaps the two most common techniques for training such models are Variational Autoencoders (VAEs) , and Generative Adversarial Networks (GANs) BID8 .", "Both have advantages and disadvantages.", "VAEs provide a meaningful lower bound on the log likelihood that is stable under training, as well as an encoding distribution from the data into the latent.", "However, they generate blurry samples due to their objective being unable to handle deterministic decoders and tractability requiring simple priors BID12 .", "On the other hand, GANs naturally enable deterministic generative models with sharply defined samples, but their training procedure is less stable .A", "relatively new approach to training generative models has emerged based on minimizing the Optimal Transport (OT) distance BID30 ) between the generative model distribution and that of the data. The", "OT approach provides a general framework for training generative models, which promises some of the best of both GANs and VAEs. Though", "interesting first results have been given in ; BID27 ; , the OT approach to generative modelling is still nascent.Our contributions are twofold: we seek to improve generative modelling capabilities with discrete and continuous latent variables, but importantly, we seek also to establish that training generative models with OT can be significantly more effective than the traditional VAE approach.Discrete latent-variable models are critical to the endeavor of unsupervised learning because of the ubiquity of discreteness in the natural world, and hence in the datasets that describe it. However", ", they are harder to train than their continuous counterparts. This has", "been tackled in a number of ways (e.g., directly mitigating high-variance discrete samples BID6 BID19 , parametrizing discrete distributions using continuous ones BID14 BID22 BID29 , deliberate model design leveraging conjugacy ).However,", "even in the simple case where the number of mixtures is small enough that monte-carlo sampling from the discrete latent is avoidable, training can still be problematic. For example", ", in BID4 a Gaussian-mixture latent-variable model (GM-LVM) was studied, and the authors were unable to train their model on MNIST using variational inference without substantially modifying the VAE objective. What appears", "to happen is that the model quickly learns to \"hack\" the VAE objective function by collapsing the discrete latent variational distribution. This problem", "only occurs in the unsupervised setting, as are able to learn the discrete latent in the semi-supervised version of the same problem once they have labeled samples for the discrete latent to latch onto. This is discussed", "in more detail in Section 2.1.The OT approach to training generative models (in particular the Wasserstein distance, discussed in Section 2.2) induces a weaker topology on the space of distributions, enabling easier convergence of distributions than in the case of VAEs BID2 . Thus, one might conjecture", "that the OT approach would enable easier training of GM-LVMs than the VAE approach. We provide evidence that this", "is indeed the case, showing that GM-LVMs can be trained in the unsupervised setting on MNIST, and motivating further the value of the OT approach to generative modelling.", "We studied an unsupervised generative model with a mixture-of-Gaussians latent variable structure, well suited to data containing discrete classes of objects with continuous variation within each class.", "We showed that such a simple and critical class of models fails to train using the VAE framework, in the sense that it immediately learns to discard its discrete-latent structure.", "We further exposed the root cause of this phenomenon with empirical results.", "We then put to the test the abstract mathematical claim that the Wasserstein distance induces a weaker topology on the space of distributions by attempting to train the same mixture-of-Gaussians model in the WAE framework.", "We found the Wasserstein objective is successful at training this model to leverage its discrete-continuous latent structure fully.", "We provided promising results on MNIST and demonstrated the additional control available to a highly structured model with both discrete and continuous latent variables.", "We hope this motivates further study of the exciting but nascent field of Optimal Transport in generative modeling." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.25531914830207825, 0.2857142686843872, 0.2545454502105713, 0.23529411852359772, 0.23529411852359772, 0.1666666567325592, 0.2448979616165161, 0.1666666567325592, 0, 0.15686273574829102, 0.08510638028383255, 0.16326530277729034, 0.25925925374031067, 0.1666666567325592, 0.21739129722118378, 0.10256409645080566, 0.1355932205915451, 0.3396226465702057, 0.17543859779834747, 0.2978723347187042, 0.25, 0.2686567008495331, 0.1860465109348297, 0.19230768084526062, 0.2641509473323822, 0.29629629850387573, 0.1538461446762085, 0.2857142686843872, 0.2666666507720947, 0.23999999463558197, 0.13636362552642822 ]

[ "This paper shows that the Wasserstein distance objective enables the training of latent variable models with discrete latents in a case where the Variational Autoencoder objective fails to do so." ]

[ "While machine learning models achieve human-comparable performance on sequential data, exploiting structured knowledge is still a challenging problem.", "Spatio-temporal graphs have been proved to be a useful tool to abstract interaction graphs and previous works exploits carefully designed feed-forward architecture to preserve such structure.", "We argue to scale such network design to real-world problem, a model needs to automatically learn a meaningful representation of the possible relations.", "Learning such interaction structure is not trivial: on the one hand, a model has to discover the hidden relations between different problem factors in an unsupervised way; on the other hand, the mined relations have to be interpretable. \n\n", "In this paper, we propose an attention module able to project a graph sub-structure in a fixed size embedding, preserving the influence that the neighbours exert on a given vertex.", "On a comprehensive evaluation done on real-world as well as toy task, we found our model competitive against strong baselines." ]

[ 0, 0, 1, 0, 0, 0 ]

[ 0.0624999962747097, 0.21621620655059814, 0.29411762952804565, 0.1304347813129425, 0.19512194395065308, 0.060606054961681366 ]

[ "A graph neural network able to automatically learn and leverage a dynamic interactive graph structure" ]

[ "We introduce NAMSG, an adaptive first-order algorithm for training neural networks.", "The method is efficient in computation and memory, and is straightforward to implement.", "It computes the gradients at configurable remote observation points, in order to expedite the convergence by adjusting the step size for directions with different curvatures in the stochastic setting.", "It also scales the updating vector elementwise by a nonincreasing preconditioner to take the advantages of AMSGRAD.", "We analyze the convergence properties for both convex and nonconvex problems by modeling the training process as a dynamic system, and provide a strategy to select the observation factor without grid search.", "A data-dependent regret bound is proposed to guarantee the convergence in the convex setting.", "The method can further achieve a O(log(T)) regret bound for strongly convex functions.", "Experiments demonstrate that NAMSG works well in practical problems and compares favorably to popular adaptive methods, such as ADAM, NADAM, and AMSGRAD.", "Training deep neural networks (Collobert et al., 2011; Hinton et al., 2012; Amodei et al., 2016; He et al., 2016) with large datasets costs a huge amount of time and computational resources.", "Efficient optimization methods are urgently required to accelerate the training process.", "First-order optimization methods (Robbins & Monro, 1951; Polyak, 1964; Bottou, 2010; Sutskever et al., 2013; Kingma & Ba, 2015; Bottou et al., 2018) are currently the most popular for training neural networks.", "They are easy to implement since only first-order gradients are introduced as input.", "Besides, they require low computation overheads except for computing gradients, which is of the same computational complexity as just evaluating the function.", "Compared with second-order methods (Nocedal, 1980; Martens, 2010; Byrd et al., 2016) , they are more effective to handle gradient noise.", "Sutskever et al. (2013) show that the momentum is crucial to improve the performance of SGD.", "Momentum methods, such as HB Polyak (1964) , can amplify steps in low-curvature eigen-directions of the Hessian through accumulation, although careful tuning is required to ensure fine convergence along the high-curvature directions.", "Sutskever et al. (2013) also rewrite the Nesterov's Accelerated Gradient (NAG) (Nesterov, 1983) in a momentum form, and show the performance improvement over HB.", "The method computes the gradient at a observation point ahead of the current point along the last updating direction.", "They illustrate that NAG suppresses the step along high curvature eigen-directions in order to prevent oscillations.", "However, all these approaches are approximation of their original forms derived for exact gradients, without fully study on gradient noise.", "show the insufficiency of HB and NAG in stochastic optimization, especially for small minibatches.", "They further present ASGD and show significant improvements.", "However, the method requires tuning of 3 parameters, leading to huge costs that impede its practical applications.", "Among variants of SGD methods, adaptive methods that scale the gradient elementwise by some form of averaging of the past gradients are particularly successful.", "ADAGRAD (Duchi et al., 2011) is the first popular method in this line.", "It is well-suited for sparse gradients since it uses all the past gradients to scale the update.", "Nevertheless, it suffers from rapid decay of step sizes, in cases of nonconvex loss functions or dense gradients.", "Subsequent adaptive methods, such as RMSPROP (Tieleman & Hinton., 2012) , ADADELTA (Zeiler, 2012) , ADAM (Kingma & Ba, 2015) , and NADAM (Dozat, 2016) , mitigate this problem by using the exponential moving averages of squared past gradients.", "However, Reddi et al. (2018) show that ADAM does not converge to optimal solutions in some convex problems, and the analysis extends to RMSPROP, ADADELTA, and NADAM.", "They propose AMSGRAD, which fixes the problem and shows improvements in experiments.", "In this paper, we propose NAMSG, that is an efficient first-order method for training neural networks.", "The name is derived from combining a configurable NAG method (CNAG) and AMSGRAD.", "NAMSG computes the stochastic gradients at configurable observation points ahead of the current parameters along the last updating direction.", "Nevertheless, instead of approximating NAG for exact gradients, it adjusts the learning rates for eigen-directions with different curvatures to expedite convergence in the stochastic setting, by selecting the observation distance.", "It also scales the update vector elementwisely using the nonincreasing preconditioner of AMSGRAD.", "We analyze the convergence properties by modeling the training process as a dynamic system, reveal the benefits of remote gradient observations and provide a strategy to select the observation factor without grid search.", "A regret bound is introduced in the convex setting, and it is further improved for strongly convex functions.", "Finally, we present experiments to demonstrate the efficiency of NAMSG in real problems.", "2 THE NAMSG SCHEME Before further description, we introduce the notations following Reddi et al. (2018) , with slight abuse of notation.", "The letter t denotes iteration number, d denotes the dimension of vectors and matrices, denotes a predefined positive small value, and S d + denotes the set of all positive definite d × d matrix.", "For a vector a ∈ R d and a matrices M ∈ R d × R d , we use a/M to denote M −1 a, diag(a) to denote a square diagonal matrix with the elements of a on the main diagonal, M i to denote the i th row of M , and", ", we use √ a for elementwise square root, a 2 for elementwise square, a/b for elementwise division, and max(a, b) to denote elementwise maximum.", "For any vector θ i ∈ R d , θ i,j denotes its j th coordinate where j ∈ {1, 2, . . . , d}.", "We define F ⊂ R d as the feasible set of points.", "Assume that F has bounded diameter D ∞ , i.e. x − y ≤ D ∞ for any x, y ∈ F, and ∇f t", "(x) ∞ ≤G ∞ , ∇f t", "(x) 1 ≤G 1 for all x ∈ F. The projection operation is defined as Π F ,A", "(y) = arg min x∈F A 1/2 (x −", "y) for A ∈ S d + and y ∈ R d .", "In the context of machine learning, we consider the minimization problem of a stochastic function,", "where x is a d dimensional vector consisting of the parameters of the model, and ξ is a random datum consisting of an input-output pair.", "Since the distribution of ξ is generally unavailable, the optimizing problem (1) is approximated by minimizing the empirical risk on the training set {ζ 1 , ζ 2 , ..., ζ N }, as", "In order to save computation and avoid overfitting, it is common to estimate the objective function and its gradient with a minibatch of training data, as", "where the minibatch S t ⊂ {1, 2, ..., N }, and b = |S t | is the size of S t .", "Firstly, we propose a configurable NAG method (CNAG).", "Since the updating directions are partially maintained in momentum methods, gradients computed at observation points, which lie ahead of the current point along the last updating direction, contain the predictive information of the forthcoming update.", "The remote observation points are defined asẋ t = x t − η t u t−1 where u t−1 is the updating vector, andẋ 1 = x 1 .", "By computing the gradient at a configurable observation pointẋ t , and substituting the gradient with the observation gradient in the HB update, we obtain the original form of CNAG, as", "where α t , β t , η t are configurable coefficients, and m 0 = 0.", "The observation distance η t can be configured to accommodate gradient noise, instead of η t = β t in NAG (Sutskever et al., 2013) .", "Both x t andẋ t are required to update in (4).", "To make the method more efficient, we simplify the update by approximation.", "Assume that the coefficients α t , β 1t , and η t , change very slowly between adjacent iterations.", "Substituting x t byẋ t + η t−1 α t−1 m t−1 , we obtain the concise form of CNAG, as", "where the observation factor µ t = η t (1 − β t )/β t , and we use x instead ofẋ for simplicity.", "In practical computation of CNAG, we further rearrange the update form as", "where only 3 scalar vector multiplications and 3 vector additions are required per iteration besides the gradient computation.", "Hereinafter, we still use (5) for simplicity in expressions.", "Then, we study the relation of CNAG and ASGD, that guides the selection of the momentum coefficient.", "shows that ASGD improves on SGD in any information-theoretically admissible regime.", "By taking a long step as well as short step and an appropriate average of both of them, ASGD tries to make similar progress on different eigen-directions.", "It takes 3 hyper-parameters: short stepα, long step parameterκ, and statistical advantage parameterξ.α is the same as the step size in SGD.", "For convex functions,κ is an estimation of the condition number.", "The statistical advantage parameterξ ≤ √κ captures trade off between statistical and computational condition numbers, andξ √κ in high stochasticity regimes.", "These hyper-parameters vary in large ranges, and are difficult to estimate.", "The huge costs in tuning limits the application of ASGD.", "The appendix shows that CNAG is a more efficient equivalent form of ASGD.", "For CNAG with constant hyper-parameters, the momentum coefficient β t = β = (κ − 0.49ξ)/(κ + 0.7ξ).", "Since the condition number is generally large in real high dimension problems, and the statistical advantage parameterξ ≤ √κ , β is close to 1.", "To sum up, the equivalence of CNAG and ASGD shows that in order to narrow the gap between the step sizes on eigen-directions with different curvatures, the momentum coefficient β should be close to 1.", "Finally, we form NAMSG by equipping CNAG with the nonincreasing preconditioner of AMSGRAD, and project the parameter vector x into the feasible set F. Algorithm 1 shows the pseudo code of NAMSG.", "Compared with AMSGRAD, NAMSG requires low computation overheads, as a scalar vector multiplication and a vector addiction per iteration, which are much cheaper than the gradient estimation.", "Almost no more memory is needed if the vector operations are run by pipelines.", "In most cases, especially when weight decay is applied for regularization, which limits the norm of the parameter vectors, the projection can also be omitted in implementation to save computation.", "We present the NAMSG method, which computes the gradients at configurable remote observation points, and scales the update vector elementwise by a nonincreasing preconditioner.", "It is efficient in computation and memory, and is straightforward to implement.", "A data-dependent regret bound is proposed to guarantee the convergence in the convex setting.", "The bound is further improved to O(log(T )) for strongly convex functions.", "The analysis of the optimizing process provides a hyperparameter policy (OBSB) which leaves only the step size for grid search.", "Numerical experiments demonstrate that NAMSG and OBSB converge faster than ADAM, NADAM, and AMSGRAD, for the tested problems." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.47999998927116394, 0.07999999821186066, 0.10256409645080566, 0.06666666269302368, 0.1428571343421936, 0.14814814925193787, 0.07407406717538834, 0.34285715222358704, 0.09756097197532654, 0.23999999463558197, 0.27272728085517883, 0.07692307233810425, 0.05714285373687744, 0.11428570747375488, 0.13793103396892548, 0.04444444179534912, 0, 0, 0.13333332538604736, 0.05882352590560913, 0.0714285671710968, 0, 0.12903225421905518, 0.17142856121063232, 0.0714285671710968, 0.13793103396892548, 0, 0.0416666641831398, 0.10256409645080566, 0, 0.3333333134651184, 0, 0, 0.09756097197532654, 0, 0.09302324801683426, 0.13333332538604736, 0.07407406717538834, 0, 0, 0.043478257954120636, 0.12121211737394333, 0, 0, 0.10810810327529907, 0, 0.06666666269302368, 0.08695651590824127, 0.1666666567325592, 0, 0, 0.04878048226237297, 0.10526315122842789, 0, 0, 0, 0, 0, 0, 0.05405404791235924, 0.0833333283662796, 0, 0.06451612710952759, 0, 0.05714285373687744, 0, 0, 0.08695651590824127, 0.0714285671710968, 0.07999999821186066, 0.052631575614213943, 0, 0, 0, 0.07999999821186066, 0, 0.07407406717538834, 0, 0.05405404791235924, 0.08888888359069824, 0, 0, 0, 0.0952380895614624, 0, 0.0833333283662796, 0.14814814925193787, 0.1538461446762085, 0.060606054961681366, 0.12903225421905518 ]

[ "A new algorithm for training neural networks that compares favorably to popular adaptive methods." ]

[ "Recent advances in Generative Adversarial Networks facilitated by improvements to the framework and successful application to various problems has resulted in extensions to multiple domains.", "IRGAN attempts to leverage the framework for Information-Retrieval (IR), a task that can be described as modeling the correct conditional probability distribution p(d|q) over the documents (d), given the query (q).", "The work that proposes IRGAN claims that optimizing their minimax loss function will result in a generator which can learn the distribution, but their setup and baseline term steer the model away from an exact adversarial formulation, and this work attempts to point out certain inaccuracies in their formulation.", "Analyzing their loss curves gives insight into possible mistakes in the loss functions and better performance can be obtained by using the co-training like setup we propose, where two models are trained in a co-operative rather than an adversarial fashion.", "Information-Retrieval (IR) involves providing a list of ranked documents {d 1 , d 2 , . . . , d k } in answer to a query q.", "This general formulation can be extended to various tasks like web-search, where the documents are web pages and information needs are queries, content-recommendation, where the documents are items/content to suggest and queries are users, and Question-Answering, where the documents are answers and queries are questions.", "The retrieved list can also be viewed as a probability distribution over candidates, one example being DISPLAYFORM0 where l is a hyperparameter.", "Even if the probability distribution is not explicit, it is desirable to retrieve a higher ranked document more often than a lower ranked document.GANs were proposed as alternatives to generative models and have been shown to be capable of modeling the true data well.", "High dimensional settings like images and word sequences have seen some success.", "Given that the generator in GANs tries to model the training data's distribution, adversarial setups seem like a natural fit for IR.", "The learned distribution can then be used to retrieve relevant documents for incoming queries.", "IRGAN is a framework proposed by , with the hope of giving Information-Retrieval, access to the large literature of GANs.IRGAN consists of a discriminator and a generator.", "Like in a typical setup, the discriminator learns to distinguish between documents produces by the real probability distribution or the real ranking and the generator's probability distribution.", "It increases the likelihood of the former and decreases it for the latter.", "The generator tries to bring its probability distribution closer to the real one so that it increases the likelihood of confusing the discriminator into believing that it is the true distribution.", "Ideally, equilibrium is achieved when the generator manages to rank the documents according to the true distribution.However, the formulation and implementation of the loss function in the work seems to have a few issues.", "Specifically, the use of the baseline term recommended in the work results in pitting the loss functions of the discriminator and the generator directly against each other and this leads to issues that are conspicuous in the loss curves.", "The training starts off with a pre-trained discriminator and generator, and the performance of the generator decreases as the training proceeds, while you would actually expect the opposite.", "When pre-training is not used, the generator does not learn at all.", "This forces IRGAN to choose the generator or discriminator based on whichever has better performance, while it expected that the generator is chosen at equilibrium.Given the traction this paper has received since its inception (53 citations as of 27 th September 2018), it is important to critically analyze the work and attribute the claimed performance improvements correctly.", "To this end, we propose two models which outperform IRGAN on two of the three tasks and give a comparable performance on the third.", "They also serve as an ablation study by experimentally showing that the generator might not be playing a vital role during train or test time.The following contributions are made in this work• We propose a model motivated by Co-training which outperforms IRGANs • We point out inaccuracies in the minimax loss function used in IRGANs • We substantiate the same by drawing conclusions from the loss curves 2 RELATED WORK 2.1 GENERATIVE ADVERSARIAL NETWORKS Generative Adversarial Networks (GANs) BID11 ) were proposed as an alternative to generative models BID23 ) which used Markov Chains or other approximations to compute intractable probability distributions.", "In essence, the generator tries to model the real data distribution and the discriminator learns to differentiate between real data points and generated data points.", "GANs are notoriously unstable to train and works like DCGANs BID21 ) and Wasserstein GAN BID1 ) have successfully attempted to alleviate a few issues.", "Nonetheless, GANs have been widely applied to various problems like image generation, text generation, cross-modal retrieval and more niche ones like Interactive Image Generation BID29 ), Text to Image ), Image to Image style transfer BID12 ) and robotics BID4 ).While", "GANs allow generation based on a random variable z, Conditional GANs BID17 ) partition the sample variable into two parts (z and y). y is", "used to denote which part of the probability distribution the generator has to generate from, and z plays the same role played in Vanilla GANs BID11 ). Conditional", "GANs dovetail with IR because y can be used to represent the query or its embedding, and in theory, the model should be able to generate the required document. DISPLAYFORM1", "We feel that an eventual adversarial formulation for IR will be similar to this in flavor.", "The experiments performed show that IRGAN is by no means state-of-the-art on those datasets.", "Further, the performance does not justify the large training time of 4 hours per generator epoch and 1 hour of discriminator epoch as opposed to 2 hours per epoch of the co-training model (11 GB GPU and Question Answering task).", "The shaky mathematical formulation renders the generator useless after training, and any gains in performance can be attributed directly to the first term of J D , where the likelihood of the real data is increased.", "We showed that the discriminator and generator are optimizing directly opposite loss functions and this is the cause of deleterious training.The poor performance of IRGAN on Web-Search and Question Answering and only a satisfactory performance on Content-Recommendation (which has dense rewards) lead us to speculate that it does not work well in sparse reward scenarios.", "This is similar to a well-known problem called the Sparse Reward Reinforcement Learning.", "We think that a correct formulation along with established techniques from the former, like reward shaping BID18 ) may lead to better performance.", "Newer methods like Hindsight Experience Replay BID0 ) which allow models to learn both from mistakes and rewards may further ameliorate learning.We would also like to explore in the direction of learning correct adversarial frameworks for more complex tasks like Image Retrieval and Question Answering which will involve learning end-toend trainable models.", "With advances in modeling sequences, this could also involve generation of documents rather than sampling them." ]

[ 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.16326530277729034, 0.1090909019112587, 0.2028985470533371, 0.21875, 0.08163265138864517, 0, 0.0833333283662796, 0.0923076868057251, 0, 0.1666666567325592, 0.09756097197532654, 0.20408162474632263, 0.1249999925494194, 0.052631575614213943, 0.039215680211782455, 0.1818181723356247, 0.072727270424366, 0.07999999821186066, 0.052631575614213943, 0.07894736528396606, 0.1249999925494194, 0.20000000298023224, 0.045454539358615875, 0.08163265138864517, 0.03389830142259598, 0.07999999821186066, 0.11538460850715637, 0.11320754140615463, 0.09302324801683426, 0.09756097197532654, 0.13793103396892548, 0.10169491171836853, 0.13333332538604736, 0.09999999403953552, 0.11999999731779099, 0.1111111044883728, 0.04651162400841713 ]

[ "Points out problems in loss function used in IRGAN, a recently proposed GAN framework for Information Retrieval. Further, a model motivated by co-training is proposed, which achieves better performance." ]

[ "Collaborative personalization, such as through learned user representations (embeddings), can improve the prediction accuracy of neural-network-based models significantly.", "We propose Federated User Representation Learning (FURL), a simple, scalable, privacy-preserving and resource-efficient way to utilize existing neural personalization techniques in the Federated Learning (FL) setting.", "FURL divides model parameters into federated and private parameters.", "Private parameters, such as private user embeddings, are trained locally, but unlike federated parameters, they are not transferred to or averaged on the server.", "We show theoretically that this parameter split does not affect training for most model personalization approaches.", "Storing user embeddings locally not only preserves user privacy, but also improves memory locality of personalization compared to on-server training.", "We evaluate FURL on two datasets, demonstrating a significant improvement in model quality with 8% and 51% performance increases, and approximately the same level of performance as centralized training with only 0% and 4% reductions.", "Furthermore, we show that user embeddings learned in FL and the centralized setting have a very similar structure, indicating that FURL can learn collaboratively through the shared parameters while preserving user privacy.", "Collaborative personalization, like learning user embeddings jointly with the task, is a powerful way to improve accuracy of neural-network-based models by adapting the model to each user's behavior (Grbovic & Cheng, 2018; Ni et al., 2018; Lee et al., 2017; Jaech & Ostendorf, 2018; McGraw et al., 2016; Vosecky et al., 2014 ).", "However, model personalization usually assumes the availability of user data on a centralized server.", "To protect user privacy, it is desirable to train personalized models in a privacy-preserving way, for example, using Federated Learning (McMahan et al., 2016; Konen et al., 2016b) .", "Personalization in FL poses many challenges due to its distributed nature, high communication costs, and privacy constraints (Li et al., 2019a; Bonawitz et al., 2019; Li et al., 2019b; 2018; Yang et al., 2019; Konen et al., 2016a) .", "To overcome these difficulties, we propose a simple, communication-efficient, scalable, privacypreserving scheme, called FURL, to extend existing neural-network personalization to FL.", "FURL can personalize models in FL by learning task-specific user representations (i.e., embeddings) (Lerer et al., 2019; Grbovic & Cheng, 2018; Ni et al., 2018; Lee et al., 2017; Jaech & Ostendorf, 2018) or by personalizing model weights (Tang & Wang, 2018) .", "Research on collaborative personalization in FL (Smith et al., 2017; Sebastian Caldas, 2019; Yao et al., 2019) has generally focused on the development of new techniques tailored to the FL setting.", "We show that most existing neural-network personalization techniques, which satisfy the split-personalization constraint (1,2,3), can be used directly in FL, with only a small change to Federated Averaging (McMahan et al., 2016) , the most common FL training algorithm.", "Existing techniques do not efficiently train user embeddings in FL since the standard Federated Averaging algorithm (McMahan et al., 2016) transfers and averages all parameters on a central server.", "Conventional training assumes that all user embeddings are part of the same model.", "Transferring all user embeddings to devices during FL training is prohibitively resource-expensive (in terms of communication and storage on user devices) and does not preserve user privacy.", "FURL defines the concepts of federated and private parameters: the latter remain on the user device instead of being transferred to the server.", "Specifically, we use a private user embedding vector on each device and train it jointly with the global model.", "These embeddings are never transferred back to the server.", "We show theoretically and empirically that splitting model parameters as in FURL affects neither model performance nor the inherent structure in learned user embeddings.", "While global model aggregation time in FURL increases linearly in the number of users, this is a significant reduction compared with other approaches (Smith et al., 2017; Sebastian Caldas, 2019) whose global aggregation time increases quadratically in the number of users.", "FURL has advantages over conventional on-server training since it exploits the fact that models are already distributed across users.", "There is little resource overhead in distributing the embedding table across users as well.", "Using a distributed embeddings table improves the memory locality of both training embeddings and using them for inference, compared to on-server training with a centralized and potentially very large user embedding table.", "Our evaluation of document classification tasks on two real-world datasets shows that FURL has similar performance to the server-only approach while preserving user privacy.", "Learning user embeddings improves the performance significantly in both server training and FL.", "Moreover, user representations learned in FL have a similar structure to those learned in a central server, indicating that embeddings are learned independently yet collaboratively in FL.", "In this paper, we make the following contributions:", "• We propose FURL, a simple, scalable, resource-efficient, and privacy preserving method that enables existing collaborative personalization techniques to work in the FL setting with only minimal changes by splitting the model into federated and private parameters.", "• We provide formal constraints under which the parameter splitting does not affect model performance.", "Most model personalization approaches satisfy these constraints when trained using Federated Averaging (McMahan et al., 2016) , the most popular FL algorithm.", "• We show empirically that FURL significantly improves the performance of models in the FL setting.", "The improvements are 8% and 51% on two real-world datasets.", "We also show that performance in the FL setting closely matches the centralized training with small reductions of only 0% and 4% on the datasets.", "• Finally, we analyze user embeddings learned in FL and compare with the user representations learned in centralized training, showing that both user representations have similar structures.", "This paper proposes FURL, a simple, scalable, bandwidth-efficient technique for model personalization in FL.", "FURL improves performance over non-personalized models and achieves similar performance to centralized personalized model while preserving user privacy.", "Moreover, representations learned in both server training and FL show similar structures.", "In future, we would like to evaluate FURL on other datasets and models, learn user embeddings jointly across multiple tasks, address the cold start problem and personalize for users not participating in global FL aggregation." ]

[ 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.0476190410554409, 0.9583333134651184, 0.0624999962747097, 0.08695651590824127, 0.09999999403953552, 0.09302324801683426, 0.1818181723356247, 0.18867923319339752, 0.11940298229455948, 0.15789473056793213, 0.23529411852359772, 0.1111111044883728, 0.3181818127632141, 0.03389830142259598, 0.23529411852359772, 0.26229506731033325, 0.2222222238779068, 0.05405404791235924, 0.0833333283662796, 0.1395348757505417, 0.1395348757505417, 0.12121211737394333, 0.17391303181648254, 0.10526315122842789, 0.04651162400841713, 0.10526315122842789, 0.15686273574829102, 0.0833333283662796, 0.21621620655059814, 0.13333332538604736, 0.0624999962747097, 0.4406779706478119, 0.10256409645080566, 0.1304347813129425, 0.20512820780277252, 0.05882352590560913, 0.21276594698429108, 0.1304347813129425, 0.31578946113586426, 0.09756097197532654, 0.1111111044883728, 0.13793103396892548 ]

[ "We propose Federated User Representation Learning (FURL), a simple, scalable, privacy-preserving and bandwidth-efficient way to utilize existing neural personalization techniques in the Federated Learning (FL) setting." ]

[ "Recurrent models for sequences have been recently successful at many tasks, especially for language modeling\n", "and machine translation.", "Nevertheless, it remains challenging to extract good representations from\n", "these models.", "For instance, even though language has a clear hierarchical structure going from characters\n", "through words to sentences, it is not apparent in current language models.\n", "We propose to improve the representation in sequence models by\n", "augmenting current approaches with an autoencoder that is forced to compress\n", "the sequence through an intermediate discrete latent space.", "In order to propagate gradients\n", "though this discrete representation we introduce an improved semantic hashing technique.\n", "We show that this technique performs well on a newly proposed quantitative efficiency measure.\n", "We also analyze latent codes produced by the model showing how they correspond to\n", "words and phrases.", "Finally, we present an application of the autoencoder-augmented\n", "model to generating diverse translations.", "Autoencoders have a long history in deep learning BID3 BID10 BID16 BID7 .", "In most cases, autoencoders operate on continuous representations, either by simply making a bottleneck BID3 , denoising BID16 , or adding a variational component BID7 .", "In many cases though, a discrete latent representation is potentially a better fit.Language is inherently discrete, and autoregressive models based on sequences of discrete symbols yield impressive results.", "A discrete representation can be fed into a reasoning or planning system or act as a bridge towards any other part of a larger system.", "Even in reinforcement learning where action spaces are naturally continuous, Metz et al. (2017) show that discretizing them and using autoregressive models can yield improvements.Unluckily, using discrete latent variables is challenging in deep learning.", "And even with continuous autoencoders, the interactions with an autoregressive component cause difficulties.", "Despite some success BID1 BID17 , the task of meaningfully autoencoding text in the presence of an autoregressive decoder has remained a challenge.In this work we present an architecture that autoencodes a sequence s of N discrete symbols from any vocabulary (e.g., a tokenized sentence), into a K-fold (we test K = 8 and K = 32) compressed sequence c(s) of Since gradient signals can vanish when propagating over discrete variables, the compression function c(s) can be hard to train.", "To solve this problem, we draw from the old technique of semantic hashing BID11 .", "There, to discretize a dense vector v one computes σ(v + n) where σ is the sigmoid function and n represents annealed Gaussian noise that pushes the network to not use middle values in v. We enhance this method by using a saturating sigmoid and a straight-through pass with only bits passed forward.", "These techniques, described in detail below, allow to forgo the annealing of the noise and provide a stable discretization mechanism that requires neither annealing nor additional loss factors.We test our discretization technique by amending language models over s with the autoencoded sequence c(s).", "We compare the perplexity achieved on s with and without the c(s) component, and contrast this value with the number of bits used in c(s).", "We argue that this number is a proper measure for the performance of a discrete autoencoder.", "It is easy to compute and captures the performance of the autoencoding part of the model.", "This quantitative measure allows us to compare the technique we introduce with other methods, and we show that it performs better than a GumbelSoftmax BID4 BID8 in this context.Finally, we discuss the use of adding the autoencoded part c(s) to a sequence model.", "We present samples from a character-level language model and show that the latent symbols correspond to words and phrases when the architecture of c(s) is local.", "ehen, we introduce a decoding method in which c(s) is sampled and then s is decoded using beam search.", "This method alleviates a number of problems observed with beam search or pure sampling.", "We show how our decoding method can be used to obtain diverse translations of a sentence from a neural machine translation model.", "To summarize, the main contributions of this paper are:(1) a discretization technique that works well without any extra losses or parameters to tune, (2) a way to measure performance of autoencoders for sequence models with baselines, (3) an improved way to sample from sequence models trained with an autoencoder part.", "In this work, the study of text autoencoders BID1 BID17 is combined with the research on discrete autoencoders BID4 BID8 .", "It turns out that the semantic hashing technique BID11 can be improved and then yields good results in this context.", "We introduce a measure of efficiency of discrete autoencoders in sequence models and show that improved semantic hashing has over 50% efficiency.", "In some cases, we can decipher the latent code, showing that latent symbols correspond to words and phrases.", "On the practical side, sampling from the latent code and then running beam search allows to get valid but highly diverse samples, an important problem with beam search BID15 .We", "leave a number of questions open for future work. How", "does the architecture of the function c(s) affect the latent code? How", "can we further improve discrete sequence autoencoding efficiency? Despite", "remaining questions, we can already see potential applications of discrete sequence autoencoders. One is", "the training of multi-scale generative models end-to-end, opening a way to generating truly realistic images, audio and video. Another", "application is in reinforcement learning. Using latent", "code may allow the agents to plan in larger time scales and explore more efficiently by sampling from high-level latent actions instead of just atomic moves." ]

[ 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.07999999821186066, 0, 0, 0.0833333283662796, 0, 0, 0.09090908616781235, 0.31578946113586426, 0, 0.08695651590824127, 0.07692307233810425, 0.07999999821186066, 0, 0, 0, 0.17391303181648254, 0.05882352590560913, 0.1621621549129486, 0.1249999925494194, 0.1395348757505417, 0.08695651590824127, 0.07692307233810425, 0, 0.14035087823867798, 0.07843136787414551, 0.06451612710952759, 0.23076923191547394, 0, 0.08163265138864517, 0.11428570747375488, 0.20689654350280762, 0.23999999463558197, 0.1249999925494194, 0.11538460850715637, 0.20689654350280762, 0, 0.12903225421905518, 0.0714285671710968, 0.10526315122842789, 0.1904761791229248, 0.0952380895614624, 0.09999999403953552, 0.07999999821186066, 0.06666666269302368, 0.1111111044883728, 0.054054051637649536 ]

[ "Autoencoders for text with a new method for using discrete latent space." ]

[ " Auto-encoding and generative models have made tremendous successes in image and signal representation learning and generation.", "These models, however, generally employ the full Euclidean space or a bounded subset (such as $[0,1]^l$) as the latent space, whose trivial geometry is often too simplistic to meaningfully reflect the structure of the data.", "This paper aims at exploring a nontrivial geometric structure of the latent space for better data representation.", "Inspired by differential geometry, we propose \\textbf{Chart Auto-Encoder (CAE)}, which captures the manifold structure of the data with multiple charts and transition functions among them.", "CAE translates the mathematical definition of manifold through parameterizing the entire data set as a collection of overlapping charts, creating local latent representations.", "These representations are an enhancement of the single-charted latent space commonly employed in auto-encoding models, as they reflect the intrinsic structure of the manifold. ", "Therefore, CAE achieves a more accurate approximation of data and generates realistic new ones.", "We conduct experiments with synthetic and real-life data to demonstrate the effectiveness of the proposed CAE.", "Autoencoding (Bourlard & Kamp, 1988; Hinton & Zemel, 1994; Liou et al., 2014 ) is a central tool in unsupervised representation learning.", "The latent space therein captures the essential information of a given data set, serving the purposes of dimension reduction, denoising, and generative modeling.", "Even for models such as generative adversarial networks (Goodfellow et al., 2014) that do not employ an encoder, the generative component starts with a latent space.", "A common practice is to model the latent space as a low-dimensional Euclidean space R l or a bounded subset of it (e.g., [0, 1] l ), sometimes equipped with a prior probability distribution.", "Such spaces carry far simple geometry and may not be adequate for representing complexly structured data.", "In this work, we are concerned with a widely studied structure: manifold.", "A commonly held belief, known as the Manifold Hypothesis (Belkin & Niyogi, 2003; Fefferman et al., 2016) , states that real-world data often lies on, or at least near, some low-dimensional manifold embedded in the high-dimensional ambient space.", "Hence, a natural approach to representation learning is to introduce a low-dimensional latent space to which the data is mapped.", "It is desirable that such a mapping possesses basic properties such as invertibility and continuity.", "In differential geometry, such a notion is coined homeomorphism.", "Challengingly, it is known that even for some simple manifolds, there does not always exist a homeomorphic mapping to the Euclidean space of the intrinsic dimension of the data.", "We elaborate two examples such examples next.", "Consider a data set X lying on the 2-dimensional sphere S 2 embedded in the ambient space R n where n ≥ 3.", "It is well known that there exist no homeomorphic maps between S 2 and an open domain on R 2 (Rotman, 2013) .", "Therefore, it is impossible for a traditional autoencoder with a 2-dimensional latent space to faithfully capture the structure of the data.", "Consequently, the dimension of the latent space needs be increased beyond the intrinsic dimension (two in this case).", "For another example, we show in Figure 1 a double torus.", "When one uses an autoencoder to map uniform data points on this manifold to R 2 , the distribution of the points is distorted and the shape destroyed, whereas if one maps to R 3 , some of the points depart from the mass and become outliers.", "Fur- thermore, in the appendix (see Figure 11 ) we demonstrate that increasing the number of parameters of the autoencoder does not help overcome the coverage issue when the latent space is a single 2-dimensional space.", "To better reflect structures, in this work, we follow the definition of manifolds in differential geometry and propose Chart Auto-Encoder (CAE) for learning a low-dimensional representation of data lying on a manifold.", "Rather than using a single function mapping, the manifold is parameterized by a collection of overlapping chart functions, each of which describes a local neighborhood and they collectively cover the entire manifold.", "To the right of Figure 1 , we show the same double torus aforementioned, now encoded by using four color-coded charts.", "One sees that the encoding result faithfully preserves the shape of the data set, as well as the topology (two holes).", "To realize the parameterization, we develop a neural network architecture and propose a training regime to implement it.", "We conduct a comprehensive set of experiments on both synethic data and real-world data to demonstrate that CAE captures much better the structure of the data and enriches the understanding of them.", "We have proposed and investigated the use of chart-based paramterization to model manifold structured data, through introducing multiple-chart latent spaces, along with transition functions, to autoencoders.", "The parameterization allows us to significantly reduce the dimension of latent encoding for efficiently representing data with complex structures.", "Numerically, we design geometric examples to analyze the behavior of the proposed CAE and illustrate its advantage over single-chart Figure 7 : Summary of benchmark test on Sphere, Genus-3, MNIST and SVHN data sets autoencoders.", "We also apply our method to real-life data sets, including MNIST and SVHN, to demonstrate the effectiveness of the proposed model.", "We believe that the proposed chart-based parameterization of manifold-structured data provides many opportunities for further analysis and applications.", "In future work, we will extend this architecture to other generative models (e.g, GAN) and apply the machinery to investigate the topology and geometry of real-world data." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.20000000298023224, 0.10810810327529907, 0.260869562625885, 0, 0.07407406717538834, 0.1428571343421936, 0, 0, 0, 0.2222222238779068, 0.3125, 0.10526315867900848, 0.09090908616781235, 0, 0.04651162400841713, 0.1818181723356247, 0, 0, 0.125, 0, 0.07407406717538834, 0, 0.23999999463558197, 0.1904761791229248, 0, 0, 0.1111111119389534, 0.05714285373687744, 0, 0, 0, 0, 0, 0.06451612710952759, 0.1599999964237213, 0, 0, 0.0833333283662796, 0.12903225421905518 ]

[ "Manifold-structured latent space for generative models" ]

[ "We tackle the problem of modeling sequential visual phenomena.", "Given examples of a phenomena that can be divided into discrete time steps, we aim to take an input from any such time and realize this input at all other time steps in the sequence.", "Furthermore, we aim to do this \\textit{without} ground-truth aligned sequences --- avoiding the difficulties needed for gathering aligned data.", "This generalizes the unpaired image-to-image problem from generating pairs to generating sequences.", "We extend cycle consistency to \\textit{loop consistency} and alleviate difficulties associated with learning in the resulting long chains of computation.", "We show competitive results compared to existing image-to-image techniques when modeling several different data sets including the Earth's seasons and aging of human faces.", "Image-to-image translation has gained tremendous attention in recent years.", "A pioneering work by shows that it is possible to realize a real image from one domain as a highly realistic and semantically meaningful image in another when paired data between the domains are available.", "Furthermore, CycleGAN extended the image-to-image translation framework in an unpaired manner by relying on the ability to build a strong prior in each domain based off generative adversarial networks (GANs, (Goodfellow et al., 2014) ) and enforcing consistency on the cyclic transformation from and to a domain.", "Methods (Kim et al., 2017; Liu et al., 2017) similar to CycleGAN have also been developed roughly around the same time.", "Since its birth, CycleGAN has become a widely adopted technique with applications even beyond computer vision (Fu et al., 2018) .", "However, CycleGAN family models are still somewhat limited since they only handle the translation problem (directly) between two domains.", "Modeling more than two domains would require separate instantiations of CycleGAN between any two pairs of domains -resulting in a quadratic model complexity.", "A major recent work, StarGAN (Choi et al., 2018) , addresses this by facilitating a fully connected domain-translation graph, allowing transformation between two arbitrary domains with a single model.", "This flexibility, however, appears restricted to domains corresponding to specific attribute changes such as emotions and appearance.", "Within nature, a multitude of settings exist where neither a set of pairs nor a fully-connected graph are the most natural representations of how one might proceed from one domain to another.", "In particular, many natural processes are sequentialand therefore the translation process should reflect this.", "A common phenomena modeled as an image-to-image task is the visual change of natural scenes between two seasons , e.g., Winter and Summer.", "This neglects the fact that nature first proceeds to Spring after Winter and Fall after Summer and therefore the pairing induces a very discontinuous reflection of the underlying process.", "Instead, we hope that by modeling a higher resolution discretization of this process, the model can more realistically approach the true model while reducing the necessary complexity of the model.", "It is difficult to obtain paired data for many image-to-image problems.", "Aligned sequential are even more difficult to come by.", "Thus, it is more plausible to gather a large number of examples from each step (domain) in a sequence without correspondences between the content of the examples.", "Therefore, we consider a setting similar to unpaired image-to-image transformation where we only have access to unaligned examples from each time step of the sequence being modeled.", "Given an example from an arbitrary point in the sequence, we then generate an aligned sequence over all other time steps -expecting a faithful realization of the image at each step.", "The key condition that required is that after generating an entire loop (returning from the last domain to the input domain), one should expect to return to the original input.", "This is quite a weak condition and promotes model flexibility.", "We denote this extension to the cycle consistency of as loop consistency and therefore name our approach as Loop-Consistent Generative Adversarial Networks (LoopGAN).", "This is a departure from many image-to-image approaches that have very short (usually length 2) paths of computation defining what it means to have gone \"there and back\", e.g. the ability to enforce reconstruction or consistency.", "Since we do not have aligned sequences, the lengths of these paths for LoopGAN are as large as the number of domains being modeled and require different approaches to make learning feasible.", "These are not entirely different from the problems that often arise in recurrent neural networks and we can draw similarities to our model as a memory-less recurrent structure with applied to images.", "We apply our method to the sequential phenomena of human aging (Zhang & Qi, 2017) and the seasons of the Alps (Anoosheh et al., 2018) with extensive comparisons with baseline methods for image-to-image translation.", "We also present additional results on gradually changing azimuth angle of chairs and gradual change of face attributes to showcased the flexibility of our model.", "We show favorable results against baseline methods for image-to-image translation in spite of allowing for them to have substantially larger model complexity.", "This is consistent with how CycleGAN is trained where two cycles are included.", "We proposed an extension to the family of image-to-image translation methods when the set of domains corresponds to a sequence of domains.", "We require that the translation task can be modeled as a consistent loop.", "This allows us to use a shared generator across all time steps leading to significant efficiency gains over a nave chaining of bi-domain image translation architectures.", "Despite this, our architecture shows favorable results when compared with the classic CycleGAN family algorithms." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.07692307233810425, 0.16326530277729034, 0.11428570747375488, 0.0714285671710968, 0.1621621549129486, 0.04878048226237297, 0.07692307233810425, 0.07999999821186066, 0.14035087823867798, 0.1621621549129486, 0.052631575614213943, 0.1111111044883728, 0.2702702581882477, 0.08888888359069824, 0.060606054961681366, 0.045454539358615875, 0, 0.04878048226237297, 0.0476190410554409, 0.04878048226237297, 0.1428571343421936, 0.23076923191547394, 0.14999999105930328, 0.1904761791229248, 0.13333332538604736, 0.04878048226237297, 0, 0.10526315122842789, 0.07692307233810425, 0.1304347813129425, 0.08510638028383255, 0.1249999925494194, 0.04999999329447746, 0.15789473056793213, 0.13793103396892548, 0.05882352590560913, 0, 0.1463414579629898, 0.0624999962747097 ]

[ "LoopGAN extends cycle length in CycleGAN to enable unaligned sequential transformation for more than two time steps." ]

[ "We propose Stochastic Weight Averaging in Parallel (SWAP), an algorithm to accelerate DNN training.", "Our algorithm uses large mini-batches to compute an approximate solution quickly and then refines it by averaging the weights of multiple models computed independently and in parallel.", "The resulting models generalize equally well as those trained with small mini-batches but are produced in a substantially shorter time.", "We demonstrate the reduction in training time and the good generalization performance of the resulting models on the computer vision datasets CIFAR10, CIFAR100, and ImageNet.", "Stochastic gradient descent (SGD) and its variants are the de-facto methods to train deep neural networks (DNNs).", "Each iteration of SGD computes an estimate of the objective's gradient by sampling a mini-batch of the available training data and computing the gradient of the loss restricted to the sampled data.", "A popular strategy to accelerate DNN training is to increase the mini-batch size together with the available computational resources.", "Larger mini-batches produce more precise gradient estimates; these allow for higher learning rates and achieve larger reductions of the training loss per iteration.", "In a distributed setting, multiple nodes can compute gradient estimates simultaneously on disjoint subsets of the mini-batch and produce a consensus estimate by averaging all estimates, with one synchronization event per iteration.", "Training with larger mini-batches requires fewer updates, thus fewer synchronization events, yielding good overall scaling behavior.", "Even though the training loss can be reduced more efficiently, there is a maximum batch size after which the resulting model tends to have worse generalization performance (McCandlish et al., 2018; Keskar et al., 2016; Hoffer et al., 2017; Golmant et al., 2018; Shallue et al., 2018) .", "This phenomenon forces practitioners to use batch sizes below those that achieve the maximum throughput and limits the usefulness of large-batch training strategies.", "Stochastic Weight Averaging (SWA) ) is a method that produces models with good generalization performance by averaging the weights of a set of models sampled from the final stages of a training run.", "As long as the models all lie in a region where the population loss is mostly convex, the average model can behave well, and in practice, it does.", "We have observed that if instead of sampling multiple models from a sequence generated by SGD, we generate multiple independent SGD sequences and average models from each, the resulting model achieves similar generalization performance.", "Furthermore, if all the independent sequences use small-batches, but start from a model trained with large-batches, the resulting model achieves generalization performance comparable with a model trained solely with small-batches.", "Using these observations, we derive Stochastic Weight Averaging in Parallel (SWAP): A simple strategy to accelerate DNN training by better utilizing available compute resources.", "Our algorithm is simple to implement, fast and produces good results with minor tuning.", "For several image classification tasks on popular computer vision datasets (CIFAR10, CIFAR100, and ImageNet), we show that SWAP achieves generalization performance comparable to models trained with small-batches but does so in time similar to that of a training run with large-batches.", "We use SWAP on some of the most efficient publicly available models to date, and show that it's able to substantially reduce their training times.", "Furthermore, we are able to beat the state of the art for CIFAR10 and train in 68% of the time of the winning entry of the DAWNBench competition." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.307692289352417, 0.10526315122842789, 0.0624999962747097, 0.1818181723356247, 0.13793103396892548, 0.17142856121063232, 0.06896550953388214, 0.17142856121063232, 0.1395348757505417, 0, 0.07999999821186066, 0.1764705777168274, 0.1538461446762085, 0.05405404791235924, 0.1395348757505417, 0.05714285373687744, 0.0555555522441864, 0.07692307233810425, 0.11999999731779099, 0.1666666567325592, 0.12121211737394333 ]

[ "We propose SWAP, a distributed algorithm for large-batch training of neural networks." ]

[ "A common way to speed up training of large convolutional networks is to add computational units.", "Training is then performed using data-parallel synchronous Stochastic Gradient Descent (SGD) with a mini-batch divided between computational units.", "With an increase in the number of nodes, the batch size grows.", "However, training with a large batch often results in lower model accuracy.", "We argue that the current recipe for large batch training (linear learning rate scaling with warm-up) is not general enough and training may diverge.", "To overcome these optimization difficulties, we propose a new training algorithm based on Layer-wise Adaptive Rate Scaling (LARS).", "Using LARS, we scaled AlexNet and ResNet-50 to a batch size of 16K.", "Training of large Convolutional Neural Networks (CNN) takes a lot of time.", "The brute-force way to speed up CNN training is to add more computational power (e.g. more GPU nodes) and train network using data-parallel Stochastic Gradient Descent, where each worker receives some chunk of global mini-batch (see e.g. BID10 or BID4 ).", "The size of a chunk should be large enough to utilize the computational resources of the worker.", "So scaling up the number of workers results in the increase of batch size.", "But using large batch may negatively impact the model accuracy, as was observed in BID10 , BID14 , BID8 , BID6 .Increasing", "the global batch while keeping the same number of epochs means that you have fewer iterations to update weights. The straight-forward", "way to compensate for a smaller number of iterations is to do larger steps by increasing the learning rate (LR). For example, BID10 suggests", "to linearly scale up LR with batch size. However using a larger LR makes", "optimization more difficult, and networks may diverge especially during the initial phase. To overcome this difficulty, BID4", "suggested a \"learning rate warm-up\": training starts with a small LR, which is slowly increased to the target \"base\" LR. With a LR warm-up and a linear scaling", "rule, BID4 successfully trained ResNet-50 BID5 ] with batch B=8K, see also BID1 ]. Linear scaling of LR with a warm-up is", "the \"state-of-the art\" recipe for large batch training.We tried to apply this linear scaling and warm-up scheme to train AlexNet BID11 ] on ImageNet BID3 ], but scaling stopped after B=2K since training diverged for large LR-s. For B=4K the accuracy dropped from the", "baseline 57.6% (B=512) to 53.1%, and for B=8K the accuracy decreased to 44.8%. To enable training with a large LR, we", "replaced Local Response Normalization layers in AlexNet with Batch Normalization (BN) BID7 ]. We will refer to this models AlexNet-BN", ". BN improved model convergence for large", "LRs as well as accuracy: for B=8K the accuracy gap decreased from 14% to 2.2%.To analyze the training stability with large", "LRs we measured the ratio between the norm of the layer weights and norm of gradients update. We observed that if this ratio is too high,", "the training becomes unstable. On other hand, if the ratio is too small, then", "weights don't change fast enough. The layer with largest ||∇W || ||W || defines", "the global limit on the learning rate. Since this ratio varies a lot between different", "layers, we can speed-up training by using a separate LR for each layer. Thus we propose a novel Layer-wise Adaptive Rate", "Scaling (LARS) algorithm.There are two notable differences between LARS and other adaptive algorithms such as ADAM BID9 ) or RMSProp BID16 ): first, LARS uses a separate learning rate for each layer and not for each weight, which leads to better stability. And second, the magnitude of the update is defined", "with respect to the weight norm for better control of training speed. With LARS we trained AlexNet-BN and ResNet-50 with", "B=16K without accuracy loss.", "Large batch is a key for scaling up training of convolutional networks.", "The existing approach for large-batch training, based on using large learning rates, leads to divergence, especially during the initial phase, even with learning rate warm-up.", "To solve these difficulties we proposed the new optimization algorithm, which adapts the learning rate for each layer (LARS) proportional to the ratio between the norm of weights and norm of gradients.", "With LARS the magnitude of the update for each layer doesn't depend on the magnitude of the gradient anymore, so it helps with vanishing and exploding gradients.", "But even with LARS and warm-up we couldn't increase LR farther for very large batches, and to keep the accuracy we have to increase the number of epochs and use extensive data augmentation to prevent over-fitting." ]

[ 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.25641024112701416, 0.0952380895614624, 0.11428570747375488, 0.2222222238779068, 0.1702127605676651, 0.5238094925880432, 0.5945945978164673, 0.17142856121063232, 0.15625, 0.20512820780277252, 0.1111111044883728, 0.13636362552642822, 0.13636362552642822, 0.12765957415103912, 0.21621620655059814, 0.04878048226237297, 0.1666666567325592, 0.1818181723356247, 0.22580644488334656, 0.25531914830207825, 0.09302324801683426, 0.06666666269302368, 0.13333332538604736, 0.13333332538604736, 0.05405404791235924, 0, 0.10526315122842789, 0.3255814015865326, 0.1690140813589096, 0.2790697515010834, 0, 0.2222222238779068, 0.2083333283662796, 0.19607841968536377, 0.1304347813129425, 0.18867923319339752 ]

[ "A new large batch training algorithm based on Layer-wise Adaptive Rate Scaling (LARS); using LARS, we scaled AlexNet and ResNet-50 to a batch of 16K." ]

[ "Finding an embedding space for a linear approximation of a nonlinear dynamical system enables efficient system identification and control synthesis.", "The Koopman operator theory lays the foundation for identifying the nonlinear-to-linear coordinate transformations with data-driven methods.", "Recently, researchers have proposed to use deep neural networks as a more expressive class of basis functions for calculating the Koopman operators.", "These approaches, however, assume a fixed dimensional state space; they are therefore not applicable to scenarios with a variable number of objects.", "In this paper, we propose to learn compositional Koopman operators, using graph neural networks to encode the state into object-centric embeddings and using a block-wise linear transition matrix to regularize the shared structure across objects.", "The learned dynamics can quickly adapt to new environments of unknown physical parameters and produce control signals to achieve a specified goal.", "Our experiments on manipulating ropes and controlling soft robots show that the proposed method has better efficiency and generalization ability than existing baselines." ]

[ 1, 0, 0, 0, 0, 0, 0 ]

[ 0.41379308700561523, 0.1538461446762085, 0.1818181723356247, 0, 0.1428571343421936, 0.1249999925494194, 0.060606054961681366 ]

[ "Learning compositional Koopman operators for efficient system identification and model-based control." ]

[ "We derive reverse-mode (or adjoint) automatic differentiation for solutions of stochastic differential equations (SDEs), allowing time-efficient and constant-memory computation of pathwise gradients, a continuous-time analogue of the reparameterization trick.\n", "Specifically, we construct a backward SDE whose solution is the gradient and provide conditions under which numerical solutions converge.\n", "We also combine our stochastic adjoint approach with a stochastic variational inference scheme for continuous-time SDE models, allowing us to learn distributions over functions using stochastic gradient descent.\n", "Our latent SDE model achieves competitive performance compared to existing approaches on time series modeling.\n", "Deterministic dynamical systems can often be modeled by ordinary differential equations (ODEs).", "For training, a memory-efficient implementation of the adjoint sensitivity method (Chen et al., 2018) effectively computes gradients through ODE solutions with constant memory cost.", "Stochastic differential equations (SDEs) are a generalization of ODEs which incorporate instantaneous noise into their dynamics (Arnold, 1974; Øksendal, 2003) .", "They are a natural fit for modeling phenomena governed by small and unobserved interactions.", "In this paper, we generalize the adjoint method to dynamics defined by SDEs resulting in an approach which we call the stochastic adjoint sensitivity method.", "Building on theoretical advances by Kunita (2019), we derive a memory-efficient adjoint method whereby we simultaneously reconstruct the original trajectory and evaluate the gradients by solving a backward SDE (in the sense of Kunita (2019)) whose formulation we detail in Section 3.", "Computationally, in order to retrace the original trajectory during the backward pass, we need to reuse noise samples generated in the forward pass.", "In Section 4, we give an algorithm that allows arbitrarily-precise querying of a Brownian motion realization at any time point, while only storing a single random seed.", "Overall, this results in a constant-memory algorithm that approximates the gradient arbitrarily well as step size reduces by computing vector-Jacobian products a constant number of times per-iteration.", "See Table 2 for a comparison of our method against previous approaches in terms of asymptotic time and memory complexity.", "We incorporate SDEs into a stochastic variational inference framework, whereby we efficiently compute likelihood ratios and backpropagate through the evidence lower bound using our adjoint approach.", "This effectively generalizes existing model families such as latent ODEs (Rubanova et al., 2019) and deep Kalman filters (Krishnan et al., 2017) ." ]

[ 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.42307692766189575, 0.27272728085517883, 0.23529411852359772, 0.09999999403953552, 0.1111111044883728, 0.3265306055545807, 0.22727271914482117, 0.15789473056793213, 0.13333332538604736, 0.20338982343673706, 0.04651162400841713, 0.07999999821186066, 0.19999998807907104, 0.2790697515010834, 0.23999999463558197, 0.08888888359069824 ]

[ "We present a constant memory gradient computation procedure through solutions of stochastic differential equations (SDEs) and apply the method for learning latent SDE models." ]

[ "It is clear that users should own and control their data and privacy.", "Utility providers are also becoming more interested in guaranteeing data privacy.", "Therefore, users and providers can and should collaborate in privacy protecting challenges, and this paper addresses this new paradigm.", "We propose a framework where the user controls what characteristics of the data they want to share (utility) and what they want to keep private (secret), without necessarily asking the utility provider to change its existing machine learning algorithms.", "We first analyze the space of privacy-preserving representations and derive natural information-theoretic bounds on the utility-privacy trade-off when disclosing a sanitized version of the data X. We present explicit learning architectures to learn privacy-preserving representations that approach this bound in a data-driven fashion.", "We describe important use-case scenarios where the utility providers are willing to collaborate with the sanitization process.", "We study space-preserving transformations where the utility provider can use the same algorithm on original and sanitized data, a critical and novel attribute to help service providers accommodate varying privacy requirements with a single set of utility algorithms.", "We illustrate this framework through the implementation of three use cases; subject-within-subject, where we tackle the problem of having a face identity detector that works only on a consenting subset of users, an important application, for example, for mobile devices activated by face recognition; gender-and-subject, where we preserve facial verification while hiding the gender attribute for users who choose to do so; and emotion-and-gender, where we hide independent variables, as is the case of hiding gender while preserving emotion detection." ]

[ 0, 0, 0, 0, 1, 0, 0, 0 ]

[ 0.09999999403953552, 0.10526315122842789, 0, 0.04999999701976776, 0.13636364042758942, 0, 0.0476190447807312, 0 ]

[ "Learning privacy-preserving transformations from data. A collaborative approach" ]

[ "Solving tasks with sparse rewards is one of the most important challenges in reinforcement learning.", "In the single-agent setting, this challenge has been addressed by introducing intrinsic rewards that motivate agents to explore unseen regions of their state spaces.", "Applying these techniques naively to the multi-agent setting results in agents exploring independently, without any coordination among themselves.", "We argue that learning in cooperative multi-agent settings can be accelerated and improved if agents coordinate with respect to what they have explored.", "In this paper we propose an approach for learning how to dynamically select between different types of intrinsic rewards which consider not just what an individual agent has explored, but all agents, such that the agents can coordinate their exploration and maximize extrinsic returns.", "Concretely, we formulate the approach as a hierarchical policy where a high-level controller selects among sets of policies trained on different types of intrinsic rewards and the low-level controllers learn the action policies of all agents under these specific rewards.", "We demonstrate the effectiveness of the proposed approach in a multi-agent gridworld domain with sparse rewards, and then show that our method scales up to more complex settings by evaluating on the VizDoom platform.", "Recent work in deep reinforcement learning effectively tackles challenging problems including the board game Go , Atari video games (Mnih et al., 2015) , and simulated robotic continuous control (Lillicrap et al., 2016) ; however, these successful approaches often rely on frequent feedback indicating whether the learning agent is performing well, otherwise known as dense rewards.", "In many tasks, dense rewards can be difficult to specify without inducing locally optimal but globally sub-optimal behavior.", "As such, it is frequently desirable to specify only a sparse reward that simply signals whether an agent has attained success or failure on a given task.", "Despite their desirability, sparse rewards introduce their own set of challenges.", "When rewards are sparse, determining which of an agent's actions led to a reward becomes more difficult, a phenomenon known in reinforcement learning as the credit-assignment problem.", "Furthermore, if rewards cannot be obtained by random actions, an agent will never receive a signal through which it can begin learning.", "As such, researchers have devised methods which attempt to provide agents with additional reward signals, known as intrinsic rewards, through which they can learn meaningful behavior (Oudeyer & Kaplan, 2009) .", "A large subset of these works focus on learning intrinsic rewards that encourage exploration of the state space (Pathak et al., 2017; Houthooft et al., 2016; Burda et al., 2019; Ostrovski et al., 2017; Tang et al., 2017) .", "Exploring the state space provides a useful inductive bias for many sparse reward problems where the challenge lies in \"finding\" rewards that may only be obtained in parts of the state space that are hard to reach by random exploration.", "These exploration-focused approaches frequently formulate their intrinsic rewards to measure the \"novelty\" of a state, such that agents are rewarded for taking actions that lead to novel states.", "Our work approaches the question of how to apply novelty-based intrinsic motivation in the cooperative multi-agent setting.", "Directly applying novelty-based intrinsic motivation to the multi-agent setting results in agents each exploring their shared state space independently from one another.", "In many cases, independent exploration may not be the most efficient method.", "For example, consider a task where multiple agents are placed in a maze and their goal is to collectively reach all of the landmarks that are spread out through the maze.", "It would be inefficient for the agents to explore the same areas redundantly.", "Instead, it would be much more sensible for agents to \"divide-and-conquer,\" or avoid redundant exploration.", "Thus, an ideal intrinsic reward for this task would encourage such behavior; however, the same behavior would not be ideal for other tasks.", "For example, take the same maze but change the task such that all agents need to reach the same landmark.", "Divide-and-conquer would no longer be an optimal exploration strategy since agents only need to find one landmark and they all need to reach the same one.", "Cooperative multi-agent reinforcement learning can benefit from sharing information about exploration across agents; however, the question of what to do with that shared information depends on the task at hand.", "In order to improve exploration in cooperative multi-agent reinforcement learning, we must first identify what kinds inductive biases can potentially be useful for multi-agent tasks and then devise intrinsic reward functions that incorporate those biases.", "Then, we must find a way to allow our agents to adapt their exploration to the given task, rather than committing to one type of intrinsic reward function.", "In this work, we first introduce a candidate set of intrinsic rewards for multiagent exploration which hold differing properties with regards to how they explore the state space.", "Subsequently, we present a hierarchical method for simultaneously learning policies trained on different intrinsic rewards and selecting the policies which maximize extrinsic returns.", "Importantly, all policies are trained using a shared replay buffer, drastically improving the sample efficiency and effectiveness of learning in cooperative multi-agent tasks with sparse rewards.", "We propose a set of multi-agent intrinsic reward functions with differing properties, and compare them both qualitatively (through videos) and quantitatively on several multi-agent exploration tasks in both a gridworld domain as well as in VizDoom.", "Overall, we can see that cooperative multi-agent tasks can, in many cases, benefit from intrinsic rewards that take into account what other agents have explored, but there are various ways to incorporate that information, each with differing properties.", "As such, we propose a method for learning policies for all intrinsic reward types simultaneously while dynamically selecting the most effective ones.", "We show that our method is capable of matching or surpassing the performance of the best performing intrinsic reward type on various tasks while using the same number of samples collected from the environment.", "In future work we hope to introduce methods for directly learning the multi-agent intrinsic reward functions, rather than selecting from a set.", "The black holes which send agents back to their starting positions if they are stepped into are an important aspect of the environment, as they add difficulty to exploration.", "The probability, ρ, of a black hole opening at each step, t, evolves as such: ρ t+1 = ρ t + N (µ, σ), where µ = σ = 0.05 for TASK 1 and µ = σ = 0.005 for 2 and 3.", "Agents observe their global position in (x, y) coordinates (scalars), as well as local information regarding walls in adjacent spaces, the probability of their adjacent spaces opening into a black hole, the relative position of other agents (if they are within 3 spaces), as well as information about which treasures the agent has already collected in the given episode.", "The global state is represented by the (x, y) coordinates of all agents, as one-hot encoded vectors for x and y separately, as well as the local information of all agents regarding black holes, walls, and treasures collected.", "Each agent's action space consists of the 4 cardinal directions as well as an option to not move, which is helpful in cases where an agent is waiting for a black hole to be safe to cross." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.0952380895614624, 0.11764705181121826, 0.17777776718139648, 0.19999998807907104, 0.2857142686843872, 0.1666666567325592, 0.3050847351551056, 0.07692307233810425, 0.04444443807005882, 0.18867923319339752, 0.054054051637649536, 0.22641508281230927, 0.08163265138864517, 0.1071428507566452, 0.1071428507566452, 0.2295081913471222, 0.18867923319339752, 0.23255813121795654, 0.20408162474632263, 0.1538461446762085, 0.18518517911434174, 0.1538461446762085, 0.1428571343421936, 0.21276594698429108, 0.09090908616781235, 0.19999998807907104, 0.145454540848732, 0.29999998211860657, 0.307692289352417, 0.2545454502105713, 0.2857142686843872, 0.18867923319339752, 0.38596490025520325, 0.1269841194152832, 0.375, 0.2142857164144516, 0.36734694242477417, 0.15094339847564697, 0.09677419066429138, 0.1111111044883728, 0.10169491171836853, 0.20338982343673706 ]

[ "We propose several intrinsic reward functions for encouraging coordinated exploration in multi-agent problems, and introduce an approach to dynamically selecting the best exploration method for a given task, online." ]

[ "Object recognition in real-world requires handling long-tailed or even open-ended data.", "An ideal visual system needs to reliably recognize the populated visual concepts and meanwhile efficiently learn about emerging new categories with a few training instances.", "Class-balanced many-shot learning and few-shot learning tackle one side of this problem, via either learning strong classifiers for populated categories or learning to learn few-shot classifiers for the tail classes.", "In this paper, we investigate the problem of generalized few-shot learning (GFSL) -- a model during the deployment is required to not only learn about \"tail\" categories with few shots, but simultaneously classify the \"head\" and \"tail\" categories.", "We propose the Classifier Synthesis Learning (CASTLE), a learning framework that learns how to synthesize calibrated few-shot classifiers in addition to the multi-class classifiers of ``head'' classes, leveraging a shared neural dictionary.", "CASTLE sheds light upon the inductive GFSL through optimizing one clean and effective GFSL learning objective.", "It demonstrates superior performances than existing GFSL algorithms and strong baselines on MiniImageNet and TieredImageNet data sets.", "More interestingly, it outperforms previous state-of-the-art methods when evaluated on standard few-shot learning.", "Visual recognition for objects in the \"long tail\" has been an important challenge to address (Wang et al., 2017; Liu et al., 2019) .", "We often have a very limited amount of data on those objects as they are infrequently observed and/or visual exemplars of them are hard to collect.", "As such, state-of-the-art methods (e.g deep learning) can not be directly applied due to their notorious demand of a large number of annotated data (Krizhevsky et al., 2017; Simonyan & Zisserman, 2014; He et al., 2016) .", "Few-shot learning (FSL) (Vinyals et al., 2016; Snell et al., 2017; Finn et al., 2017 ) is mindful of the limited instances (i.e, shots) per \"tail\" concept, which attempts to address this challenging problem by distinguishing between the data-rich \"head\" categories as SEEN classes and data-scarce \"tail\" categories as UNSEEN classes.", "While it is difficult to build classifiers with data from UNSEEN classes, FSL leverages data from SEEN classes to extract inductive biases for effective classifiers acquisition on UNSEEN ones.", "We refer to (Larochelle, 2018) for an up-to-date survey in few-shot learning.", "This type of learning, however, creates a chasm in object recognition.", "Classifiers from many-shot learning for SEEN classes and those from few-shot learning for UNSEEN classes do not mix -they cannot be combined directly to recognize all object categories at the same time.", "In this paper, we study the problem of Generalized Few-Shot Learning (GFSL), which focuses on the joint classification of both data-rich and data-poor categories.", "In particular, our goal is for the model trained on the SEEN categories to be capable of incorporating limited UNSEEN class instances, and make predictions for test instances in both the \"head\" and \"tail\" of the entire distribution of categories.", "Figure 1 illustrates the high-level idea of our proposal, contrasting the standard few-shot learning.", "In contrast to prior works (Hariharan & Girshick, 2017; Wang et al., 2017; Liu et al., 2019 ) that focus on learning \"head\" and \"tail\" concepts in a transductive manner, our learning setup requires inductive modeling of the\"tail\", which is therefore more challenging as we assume no knowledge about the UNSEEN \"tail\" categories is available during the model learning phase.", "(GFSL) .", "GFSL requires to extract inductive bias from SEEN categories to facilitate efficiently learning on few-shot UNSEEN \"tail\" categories, while maintaining discernability on \"head\" classes.", "To this end, we propose Classifier Synthesis Learning (CASTLE), where the few-shot classifiers are synthesized based on a shared neural dictionary across classes.", "Such synthesized few-shot classifiers are then used together with the many-shot classifiers.", "To this purpose, we create a scenario, via sampling a set of instances from SEEN categories and pretend that they come from UNSEEN, and apply the synthesized classifiers (based on the instances) as if they are many-shot classifiers to optimize multi-class classification together with the remaining many-shot SEEN classifiers.", "In other words, we construct few-shot classifiers to not only perform well on the few-shot classes but also to be competitive when used in conjunction with many-shot classifiers of populated classes.", "We argue that such highly contrastive learning can benefit few-shot classification with high discernibility in its learned visual embeddings (cf. Section 4.2 and Section 4.4).", "We empirically validate our approach on two standard benchmark data sets -MiniImageNet and TieredImageNet.", "The proposed approach retains competitive \"head\" concept recognition performances while outperforming existing approaches on few-shot learning and generalized few-shot learning.", "We highlight that CASTLE has learned a better calibration between many-shot SEEN classifiers and synthesized UNSEEN classifiers, which naturally addresses the confidence mismatch phenomena , i.e, SEEN and UNSEEN classifiers have different confidence ranges.", "Building a high-quality visual system usually requires to have a large scale annotated training set with many shots per categories.", "Many large-scale datasets such as ImageNet have an ample number of instances for popular classes (Russakovsky et al., 2015; Krizhevsky et al., 2017) .", "However, the data-scarce \"tail\" of the category distribution matters.", "For example, a visual search engine needs to deal with the rare object of interests (e.g endangered species) or newly defined items (e.g new smartphone models), which only possess a few data instances.", "Directly training a system over all classes is prone to over-fit and can be biased towards the data-rich categories.", "Few-shot learning (FSL) is proposed to tackle this problem, via meta-learning an inductive bias from the SEEN classes, such that it transfers to the learning process of UNSEEN classes with few training data during the model deployment.", "For example, one line of works uses meta-learned discriminative feature embeddings (Snell et al., 2017; Oreshkin et al., 2018; Rusu et al., 2018; Scott et al., 2018; Ye et al., 2018; Lee et al., 2019) together with non-parametric nearest neighbor classifiers, to recognize novel classes given a few exemplars.", "Another line of works (Finn et al., 2017; Nichol et al., 2018; Lee & Choi, 2018; Antoniou et al., 2018; Vuorio et al., 2018) chooses to learn a common initialization to a pre-specified model configuration and adapt rapidly using fixed steps of gradient descents over the few-shot training data from UNSEEN categories.", "FSL emphasizes on building models of the UNSEEN classes and ignore its real-world use case of assisting the many-shot recognition of the \"'head\" categories.", "A more realistic setting, i.e, low-shot learning, has been studied before (Hariharan & Girshick, 2017; Wang et al., 2018; Gao et al., 2018; Ye et al., 2018; Liu et al., 2019) .", "The main aim is to recognize the entire set of concepts in a transductive learning framework -during the training of the target model, you have access to both the SEEN and UNSEEN categories.", "The key difference to our proposed GFSL is that we assume no access to UNSEEN classes in the learning phase, which requires the model to inductively transfer knowledge from SEEN classes to UNSEEN ones during the evaluation.", "Previous approaches mostly focus on the transductive setup of GFSL.", "Some of them (Hariharan & Girshick, 2017; Wang et al., 2018; Gao et al., 2018) apply the exemplar-based classification paradigms on both SEEN and UNSEEN categories to resolve the transductive learning problem.", "Others (Wang et al., 2017; Schönfeld et al., 2018; Liu et al., 2019) usually ignore the explicit relationship between SEEN and UNSEEN categories, and learn separate classifiers.", "Ren et al. (2018a) ; Gidaris & Komodakis (2018) propose to solve inductive GFSL via either composing UNSEEN with SEEN classifiers or meta-leaning with recurrent back-propagation procedure.", "Gidaris & Komodakis (2018) is the most related work to CASTLE, where we differ in how we compose classifiers and the unified learning objective, i.e, we used a learned neural dictionary instead of using MC classifiers as bases.", "In summary, CASTLE learns both many-shot classifiers and synthesized classifiers via optimizing a single unified objective function, where a classifier composition model with a neural dictionary is leveraged for assembling few-shot classifiers.", "Our experiments highlight that CASTLE not only outperforms existing methods in terms of GFSL performances from many different aspects, but more interestingly, also improves the classifier's discernibility over standard FSL." ]

[ 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0, 0.14999999105930328, 0.4000000059604645, 0.1599999964237213, 0.22727271914482117, 0.25806450843811035, 0.1249999925494194, 0.06896550953388214, 0.10526315122842789, 0.09999999403953552, 0.039215680211782455, 0.06666666269302368, 0.14999999105930328, 0.2857142686843872, 0, 0.22727271914482117, 0.052631575614213943, 0.1249999925494194, 0.06896550953388214, 0.057971011847257614, 0.15789473056793213, 0.1538461446762085, 0.2222222238779068, 0.145454540848732, 0.1860465109348297, 0.1463414579629898, 0.13333332538604736, 0.11764705181121826, 0.17391303181648254, 0.05714285373687744, 0.052631575614213943, 0, 0.0416666604578495, 0.11428570747375488, 0.04081632196903229, 0.07547169178724289, 0.17241379618644714, 0.1111111044883728, 0, 0.09090908616781235, 0.08695651590824127, 0.07692307233810425, 0.08695651590824127, 0.1538461446762085, 0.1904761791229248, 0.15686273574829102, 0.3181818127632141, 0.043478257954120636 ]

[ "We propose to learn synthesizing few-shot classifiers and many-shot classifiers using one single objective function for GFSL." ]

[ "Machine learning workloads are often expensive to train, taking weeks to converge.", "The current generation of frameworks relies on custom back-ends in order to achieve efficiency, making it impractical to train models on less common hardware where no such back-ends exist.", "Knossos builds on recent work that avoids the need for hand-written libraries, instead compiles machine learning models in much the same way one would compile other kinds of software.", "In order to make the resulting code efficient, the Knossos complier directly optimises the abstract syntax tree of the program.", "However in contrast to traditional compilers that employ hand-written optimisation passes, we take a rewriting approach driven by the $A^\\star$ search algorithm and a learn value function that evaluates future potential cost reduction of taking various rewriting actions to the program.", "We show that Knossos can automatically learned optimisations that past compliers had to implement by hand.", "Furthermore, we demonstrate that Knossos can achieve wall time reduction compared to a hand-tuned compiler on a suite of machine learning programs, including basic linear algebra and convolutional networks.", "The Knossos compiler has minimal dependencies and can be used on any architecture that supports a \\Cpp toolchain. \n", "Since cost model the proposed algorithm optimises can be tailored to a particular hardware architecture, the proposed approach can potentially applied to a variety of hardware.", "While the development of any kind of software can benefit from compliers able to produce fast code, runtime efficiency is particularity important for modern machine learning.", "In particular, because modern models they can take weeks to train (OpenAI, 2018) , complier optimisations that lead to execution speed-ups are of huge value.", "In parallel, machine learning is being deployed on a variety of diverse devices ranging from wearables to huge clusters clusters of powerful GPUs.", "Since each architecture has different performance profile and requires different code optimisations, it is difficult to provide tooling that works fast on all of them.", "Traditionally, the tension between performance and interoperability is resolved by machine learning frameworks (Paszke et al., 2017; Abadi et al., 2016) .", "In these frameworks, while code execution is outsourced to hardware-specific back-ends such as XLA (XLA authors, 2016) .", "While this approach has seen huge initial success, the cost of providing customised back-ends for each target architecture is prohibitive.", "Moreover, the frameworks also custom front-ends that require the programmer to specify the model being trained as a compute graph.", "Since the compute graph has semantics separate from the host programming language, this process is often error-prone and time-consuming.", "In order to address these obstacles, a new generation of tools has recently appeared that transform machine learning code using the same techniques that have been used for compiling traditional software.", "The need for a separate front-end API for machine learning operations is eliminated by including automatic differentiation as a first-class feature of the complied language (Innes et al., 2019; Frostig et al., 2018) .", "Instead of custom back-ends, modern machine learning compliers use an intermediate representation and perform extensive code optimisations (Innes et al., 2019; Frostig et al., 2018; van Merrienboer et al., 2018; Wei et al., 2018; Sotoudeh et al., 2019; Rotem et al., 2018) .", "In addition, program optimisation is being modelled as a machine learning task itself, with the complier learning how to perform rewrites (Chen et al., 2018b; a) .", "in mind.", "We formalize program optimisation as a finite-horizon Markov Decision Process (MDP), with the reward signal determined by the cost of executing a program.", "By solving this MDP, we are able to produce fast code tailor-made for any given task and architecture, without relying on backend-specific hand-written libraries.", "Knossos works by re-writing programs written in an intermediate representation (IR).", "Akin to JAX (Frostig et al., 2018) and Zygote (Innes et al., 2019) , all Knossos functions are potentially differentiable, avoiding the syntactic awkwardness that arises from embedding a differentiable program in a host language.", "The IR can then be transpiled, allowing it to run on any platform that supports a C ++ toolchain.", "This allows Knossos code to be seamlessly deployed on specialized or embedded hardware without the need of manual tuning, both for training and for deployment of models, enabling a much broader user base than competing approaches.", "To our knowledge, Knossos is the first compiler that combines RL-based program optimisation, firstclass support for deep learning primitives and the ability to target any architecture supporting the C ++ toolchain.", "We defer detailed scope comparisons with prior work to Section 4.", "We empirically demonstrate the benefits of our program optimisation in Section 5, showing that Knossos was able to automatically learn loop fusion, a type of compiler optimisation that previously had to be applied manually.", "We have introduced Knossos, a new complier targetting machine learning and numerical computation.", "Thanks to its automatic code optimisation, Knossos produces binaries that achieve better run-times than a traditional, rule-based complier.", "Knossos can deal with complex code generated by automatic differentiation and automatically discover optimisations that previously required careful complier design.", "We believe that Knossos will pave the way towards a new generation of future compliers, which will crucially rely on automatically inferring the correct optimisations.", "It also has a LISP-like surface syntax, which we used to implement our programs.", "In the future, we plan to provide transpilers, allowing for the compilation of code written in other languages into Knossos.", "We provide a sample Knossos program in Figure 4 .In", "order to facilitate Machine Learning workloads, the Knossos IL has native support for automatic differentiation. We", "use a new unified view of automatic differentiation as generalised transposition (Elliott, 2018) . Rather", "than having an explicit distinction between forward mode and reverse mode AD, Knossos uses uses a type system together with a set of consistent rewrite rules. Whenever", "the gradient operator is used as part of a Knossos algorithm, the complier first generates a syntax tree corresponding to the differentiated program and then applies rewrites to optimize the cost of its execution. This means", "that the resulting AD algorithm is tailor-made and optimized with that exact use case in mind. This is in", "contrast to systems such as PyTorch, which have hard-coded routines for backward-mode AD. From the perspective", "of the user, this process is completely transparent in the sense that taking gradients can be applied to any piece of Knossos code.", "While the details of this process are beyond the scope of this paper, from the perspective of this work, the important feature of AD is that it corresponds to a transformation of the abstract syntax tree.", "The resulting AST can then be optimised in the same way as any other code." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.17391303181648254, 0.052631575614213943, 0.09999999403953552, 0.13793103396892548, 0.0833333283662796, 0.14814814925193787, 0.14999999105930328, 0, 0.0624999962747097, 0.1621621549129486, 0.0555555522441864, 0.1818181723356247, 0.1111111044883728, 0.1249999925494194, 0.13793103396892548, 0, 0.06666666269302368, 0, 0.1904761791229248, 0.0952380895614624, 0.1428571343421936, 0.21052631735801697, 0.1249999925494194, 0.1111111044883728, 0, 0.045454543083906174, 0.06451612710952759, 0.08695651590824127, 0.09756097197532654, 0.260869562625885, 0.0952380895614624, 0.23999999463558197, 0.13333332538604736, 0.1249999925494194, 0.05714285373687744, 0.07692307233810425, 0.12903225421905518, 0.09090908616781235, 0.1428571343421936, 0, 0.0555555522441864, 0.0476190447807312, 0.0714285671710968, 0.0714285671710968, 0.12121211737394333, 0.052631575614213943, 0.07407406717538834 ]

[ "We combine A* search with reinforcement learning to speed up machine learning code" ]

[ "Grasping an object and precisely stacking it on another is a difficult task for traditional robotic control or hand-engineered approaches.", "Here we examine the problem in simulation and provide techniques aimed at solving it via deep reinforcement learning.", "We introduce two straightforward extensions to the Deep Deterministic Policy Gradient algorithm (DDPG), which make it significantly more data-efficient and scalable.", "Our results show that by making extensive use of off-policy data and replay, it is possible to find high-performance control policies.", "Further, our results hint that it may soon be feasible to train successful stacking policies by collecting interactions on real robots.", "Dexterous manipulation is a fundamental challenge in robotics.", "Researchers have long sought a way to enable robots to robustly and flexibly interact with fixed and free objects of different shapes, materials, and surface properties in the context of a broad range of tasks and environmental conditions.", "Such flexibility is very difficult to achieve with manually designed controllers.", "The recent resurgence of neural networks and \"deep learning\" has inspired hope that these methods will be as effective in the control domain as they are for perception.", "Indeed, recent work has used neural networks to learn solutions to a variety of control problems BID31 BID6 BID30 BID10 BID16 .While", "the flexibility and generality of learning approaches is promising for robotics, these methods typically require a large amount of data that grows with the complexity of the task. What", "is feasible on a simulated system, where hundreds of millions of control steps are possible BID22 BID31 , does not necessarily transfer to real robot applications due to unrealistic learning times. One", "solution to this problem is to restrict the generality of the controller by incorporating task specific knowledge, e.g. in the form of dynamic movement primitives BID29 , or in the form of strong teaching signals, e.g. kinesthetic teaching of trajectories BID23 . Recent", "works have had success learning flexible neural network policies directly on real robots (e.g. BID4 BID38 ), but tasks as complex as precise grasping-and-stacking remain daunting.In this paper we investigate in simulation the possibility of learning precise manipulation skills endto-end with a general purpose model-free deep reinforcement learning algorithm. We assess", "the feasibility of performing analogous experiments on real robotics hardware and provide guidance with respect to the choice of learning algorithm, experimental setup, and the performance that we can hope to achieve.We consider the task of picking up a Lego brick from the table and stacking it onto a second nearby brick using a robotic arm and gripper. This task", "involves contact-rich interactions between the robotic arm and two freely moving objects. It also requires", "mastering several sub-skills (reaching, grasping, lifting, and stacking) . Each of these sub-skills", "is challenging in its own right as they require both precision (for instance, successful stacking requires accurate alignment of the two bricks) and as well as robust generalization over a large state space (e.g. different initial positions of the bricks and the initial configuration of the arm). Finally, there exist non-trivial", "and long-ranging dependencies between the solutions for different sub-tasks: for instance, the ability to successfully stack the brick depends critically on having picked up the brick in a sensible way beforehand. This paper makes several contributions", ": 1. We build on the Deep Deterministic", "Policy Gradient (DDPG; ), a general purpose model-free reinforcement learning algorithm for continuous actions, and extend it in two ways: firstly, we improve the data efficiency of the algorithm by scheduling updates of the network parameters independently of interactions with the environment. Secondly, we overcome the computational", "and experimental bottlenecks of single-machine single-robot learning by introducing a distributed version of DDPG which allows data collection and network training to be spread out over multiple computers and robots. 2. We show how to use these straightforward", "algorithmic developments to solve a complex, multi-stage manipulation problem. We further propose two broadly applicable strategies", "that allow us to reliably find solutions to complex tasks and further reduce the amount of environmental interaction. The first of these strategies is a recipe for designing", "effective shaping rewards for compositional tasks, while the second biases the distribution of initial states to achieve an effect akin a form of apprenticeship learning.In combination these contributions allow us to reliably learn robust policies for the full stacking task from scratch in less than 10 million environment transitions. This corresponds to less than 10 hours of interaction time", "on 16 robots. In addition, we show that when states from demonstration trajectories", "are used as the start states for learning trials the full task can be learned with 1 million transitions (i.e. less than 1 hour of interaction on 16 robots). To our knowledge our results provide the first demonstration of end-to-end", "learning for a complex manipulation problem involving multiple freely moving objects. They are also suggest that it may be possible to learn such non-trivial manipulation", "skills directly on real robots.", "We have introduced two extensions to the DDPG algorithm which make it a practical method for learning robust policies for complex continuous control tasks.", "We have shown that by decoupling the frequency of network updates from the environment interaction we can dramatically improve data-efficiency.", "Parallelizing data acquisition and learning substantially reduces wall clock time.In addition, we presented two methods that help to guide the learning process towards good solutions and thus reduce the pressure on exploration strategies and speed up learning.", "In combination these contributions allow us to solve a challenging manipulation problem end-to-end, suggesting that many hard control problems lie within the reach of modern learning methods.It is of course challenging to judge the transfer of results in simulation to the real world.", "We have taken care to design a physically realistic simulation, and in initial experiments, which we have performed both in simulation and on the physical robot, we generally find a good correspondence of performance and learning speed between simulation and real world.", "This makes us optimistic that performance numbers may also hold when going to the real world.", "A second limitation of our simulated setup is that it currently uses information about the state of the environment would require additional instrumentation of the experimental setup, e.g. to determine the position of the two bricks in the work space.", "These are issues that need to be addressed with care as experiments move to robotics hardware in the lab.", "Nevertheless, the algorithms and techniques presented here offer important guidance for the application of deep reinforcement learning methods to dexterous manipulation on a real robot." ]

[ 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.06451612710952759, 0.20689654350280762, 0.0624999962747097, 0.1249999925494194, 0.25, 0, 0.0476190447807312, 0.09090908616781235, 0.052631575614213943, 0.1249999925494194, 0.0555555522441864, 0.09756097197532654, 0.04651162400841713, 0.1666666567325592, 0.13793103396892548, 0, 0, 0.03703703358769417, 0.0476190447807312, 0, 0.07999999821186066, 0.13333332538604736, 0.07407406717538834, 0.054054051637649536, 0.13114753365516663, 0, 0.1702127605676651, 0.17142856121063232, 0, 0.1764705777168274, 0.06666666269302368, 0.09090908616781235, 0.0833333283662796, 0.08888888359069824, 0.07407406717538834, 0.045454543083906174, 0.13793103396892548, 0.22857142984867096 ]

[ "Data-efficient deep reinforcement learning can be used to learning precise stacking policies." ]

[ "In recent years deep reinforcement learning has been shown to be adept at solving sequential decision processes with high-dimensional state spaces such as in the Atari games.", "Many reinforcement learning problems, however, involve high-dimensional discrete action spaces as well as high-dimensional state spaces.", "In this paper, we develop a novel policy gradient methodology for the case of large multidimensional discrete action spaces.", "We propose two approaches for creating parameterized policies: LSTM parameterization and a Modified MDP (MMDP) giving rise to Feed-Forward Network (FFN) parameterization.", "Both of these approaches provide expressive models to which backpropagation can be applied for training.", "We then consider entropy bonus, which is typically added to the reward function to enhance exploration.", "In the case of high-dimensional action spaces, calculating the entropy and the gradient of the entropy requires enumerating all the actions in the action space and running forward and backpropagation for each action, which may be computationally infeasible.", "We develop several novel unbiased estimators for the entropy bonus and its gradient.", "Finally, we test our algorithms on two environments: a multi-hunter multi-rabbit grid game and a multi-agent multi-arm bandit problem.", "In recent years deep reinforcement learning has been shown to be adept at solving sequential decision processes with high-dimensional state spaces such as in the Go game ) and Atari games BID6 , BID7 , BID9 , BID11 , BID12 , BID22 , BID1 ).", "In all of these success stories, the size of the action space was relatively small.", "Many reinforcement learning problems, however, involve high-dimensional action spaces as well as high-dimensional state spaces.", "Examples include StarCraft BID21 , BID4 ), where there are many agents each of which can take a finite number of actions; and coordinating self-driving cars at an intersection, where each car can take a finite set of actions BID17 ).In", "this paper, we develop a novel policy gradient methodology for the case of large multidimensional action spaces. There", "are two major challenges in developing such a methodology:• For large multidimensional action spaces, how can we design expressive and differentiable parameterized policies which can be efficiently sampled?• In policy", "gradient, in order to encourage sufficient exploration, an entropy bonus term is typically added to the objective function. However, in", "the case of high-dimensional action spaces, calculating the entropy and its gradient requires enumerating all the actions in the action space and running forward and backpropagation for each action, which may be computationally infeasible. How can we", "efficiently approximate the entropy and its gradient while maintaining desirable exploration?In this paper", ", we first propose two approaches for parameterizing the policy: a LSTM model and a Modified MDP (MMDP) giving rise to Feed-Forward Network (FFN) model. For both of", "these parameterizations, actions can be efficiently sampled from the policy distribution, and backpropagation can be employed for training. We then develop", "several novel unbiased estimators for the entropy bonus and its gradient. These estimators", "can be combined with stochastic gradient descent giving a new a class of policy gradient algorithms with desirable exploration. Finally, we test", "our algorithms on two environments: a multi-agent multi-arm bandit problem and a multi-agent hunter-rabbit grid game.", "In this paper, we developed a novel policy gradient methodology for the case of large multidimensional discrete action spaces.", "We proposed two approaches for creating parameterized policies: LSTM parameterization and a Modified MDP (MMDP) giving rise to Feed-Forward Network (FFN) parameterization.", "Both of these approaches provide expressive models to which backpropagation can be applied for training.", "We then developed several novel unbiased estimators for entropy bonus and its gradient.", "We did experimental work for two environments with large multidimensional action space.", "For these environments, we found that both the LSTM and MMDP approach could successfully solve large multidimensional action space problems, with the LSTM approach generally performing better.", "We also found that the smoothed estimates of the entropy and the unbiased gradient estimate of the entropy gradient can help reduce computational cost while not sacrificing significant loss in performance." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0 ]

[ 0.04878048226237297, 0.14814814925193787, 0.3636363446712494, 0.17142856121063232, 0.06896550953388214, 0.06896550953388214, 0.2380952388048172, 0.37037035822868347, 0.0624999962747097, 0.07547169178724289, 0.14814814925193787, 0.07692307233810425, 0.0416666641831398, 0.3125, 0.23255813121795654, 0.0624999962747097, 0.2222222238779068, 0.14814814925193787, 0.14999999105930328, 0.1818181723356247, 0.38461539149284363, 0.12121211737394333, 0.0714285671710968, 0.3636363446712494, 0.17142856121063232, 0.06896550953388214, 0.37037035822868347, 0.4615384638309479, 0.31578946113586426, 0.1538461446762085 ]

[ "policy parameterizations and unbiased policy entropy estimators for MDP with large multidimensional discrete action space" ]

[ "Neural networks offer high-accuracy solutions to a range of problems, but are computationally costly to run in production systems.", "We propose a technique called Deep Learning Approximation to take an already-trained neural network model and build a faster (and almost equally accurate) network by manipulating the network structure and coefficients without requiring re-training or access to the training data.", "Speedup is achieved by applying a sequential series of independent optimizations that reduce the floating-point operations (FLOPs) required to perform a forward pass.", "An optimal lossy approximation is chosen for each layer by weighing the relative accuracy loss and FLOP reduction.", "On PASCAL VOC 2007 with the YOLO network, we show an end-to-end 2x speedup in a network forward pass with a $5$\\% drop in mAP that can be re-gained by finetuning, enabling this network (and others like it) to be deployed in compute-constrained systems." ]

[ 0, 0, 0, 0, 1 ]

[ 0.07692307233810425, 0.0476190447807312, 0.06666666269302368, 0, 0.08695651590824127 ]

[ "Decompose weights to use fewer FLOPs with SVD" ]

[ "Asynchronous distributed methods are a popular way to reduce the communication and synchronization costs of large-scale optimization.", "Yet, for all their success, little is known about their convergence guarantees in the challenging case of general non-smooth, non-convex objectives, beyond cases where closed-form proximal operator solutions are available.\n", "This is all the more surprising since these objectives are the ones appearing in the training of deep neural networks.\n\n", "In this paper, we introduce the first convergence analysis covering asynchronous methods in the case of general non-smooth, non-convex objectives.", "Our analysis applies to stochastic sub-gradient descent methods both with and without block variable partitioning, and both with and without momentum.", "It is phrased in the context of a general probabilistic model of asynchronous scheduling accurately adapted to modern hardware properties.", "We validate our analysis experimentally in the context of training deep neural network architectures.", "We show their overall successful asymptotic convergence as well as exploring how momentum, synchronization, and partitioning all affect performance.", "Training parameters arising in Deep Neural Net architectures is a difficult problem in several ways (Goodfellow et al., 2016) .", "First, with multiple layers and nonlinear activation functions such as sigmoid and softmax functions, the ultimate optimization problem is nonconvex.", "Second, with ReLU activation functions and max-pooling in convolutional structures, the problem is nonsmooth, i.e., it is not differentiable everywhere, although typically the set of non-differentiable points is a set of measure zero in the space of the parameters.", "Finally, in many applications it is unreasonable to load the whole sample size in memory to evaluate the objective function or (sub)gradient, thus samples must be taken, necessitating analysis in a probabilistic framework.", "The analysis of parallel optimization algorithms using shared memory architectures, motivated by applications in machine learning, was ushered in by the seminal work of Recht et al. (2011) (although precursors exist, see the references therein).", "Further work refined this analysis, e.g. (Liu & Wright, 2015) and expanded it to nonconvex problems, e.g. (Lian et al., 2015) .", "However, in all of these results, a very simplistic model of asynchronous computation is presented to analyze the problem.", "Notably, it is assumed that every block of the parameter, among the set of blocks of iterates being optimized, has a fixed, equal probability of being chosen at every iteration, with a certain vector of delays that determine how old each block is that is stored in the cache relative to the shared memory.", "As one can surmise, this implies complete symmetry with regards to cores reading and computing the different blocks.", "This does not correspond to asynchronous computation in practice.", "In particular, in the common Non-Uniform Memory Access (NUMA) setting, practical experience has shown that it can be effective for each core to control a set of blocks.", "Thus, the choice of blocks will depend on previous iterates, which core was last to update, creating probabilistic dependence between the delay vector and the choice of block.", "This exact model is formalized in Cannelli et al., which introduced a new probabilistic model of asynchronous parallel optimization and presented a coordinate-wise updating successive convex approximation algorithm.", "In this paper, we are interested in studying parallel asynchronous stochastic subgradient descent for general nonconvex nonsmooth objectives, such as the ones arising in the training of deep neural network architectures.", "Currently, there is no work in the literature specifically addressing this problem.", "The closest related work is given by Zhu et al. (2018) and , which consider asynchronous proximal gradient methods for solving problems of the form f (x) + g(x), where f is smooth and nonconvex, and g(x) is nonsmooth, with an easily computable closed form prox expression.", "This restriction applies to the case of training a neural network which has no ReLUs or max pooling in the architecture itself, i.e., every activation is a smooth function, and there is an additional regularization term, such as an 1 .", "These papers derive expected rates of convergence.", "In the general case, where the activations themselves are nonsmooth-for instance in the presence of ReLUs-there is no such additive structure, and no proximal operator exists to handle away the non-smoothness and remove the necessity of computing and using subgradients explicitly in the optimization procedure.", "This general problem of nonsmooth nonconvex optimization is already difficult (see, e.g., Bagirov et al. (2014) ), and the introduction of stochastically uncertain iterate updates creates an additional challenge.", "Classically, the framework of stochastic approximation, with stochastic estimates of the subgradient approximating elements in a differential inclusion that defines a flow towards minimization of the objective function, is a standard, effective approach to analyzing algorithms for this class of problems.", "Some texts on the framework include Kushner & Yin (2003) , which we shall reference extensively in the paper, and Borkar (2008) .", "See also Ermol'ev & Norkin (1998) and Ruszczyński (1987) for some classical results in convergence of stochastic algorithms for nonconvex nonsmooth optimization.", "Interest in stochastic approximation has resurfaced recently sparked by the popularity of Deep Neural Network architectures.", "For instance, see the analysis of nonconvex nonsmooth stochastic optimization with an eye towards such models in Davis et al. (2018) and Majewski et al. (2018) .", "In this paper, we provide the first analysis for nonsmooth nonconvex stochastic subgradient methods in a parallel asynchronous setting, in the stochastic approximation framework.", "For this, we employ the state of the art model of parallel computation introduced in Cannelli et al., which we map onto the analysis framework of Kushner & Yin (2003) .", "We prove show that the generic asynchronous stochastic subgradient methods considered are convergent, with probability 1, for nonconvex nonsmooth functions.", "This is the first result for this class of algorithms, and it combines the state of the art in these two areas, while extending the scope of the results therein.", "Furthermore, given the success of momentum methods (see, e.g., Zhang et al. (2017) ), we consider the momentum variant of the classical subgradient method, again presenting the first convergence analysis for this class of algorithms.", "We validate our analysis numerically by demonstrating the performance of asynchronous stochastic subgradient methods of different forms on the problem of ResNet deep network training.", "We shall consider variants of asynchronous updating with and without write locks and with and without block variable partitioning, showing the nuances in terms of convergence behavior as depending on these strategies and properties of the computational hardware.", "In this paper we analyzed the convergence theory of asynchronous stochastic subgradient descent.", "We found that the state of the art probabilistic model on asynchronous parallel architecture applied to the stochastic subgradient method, with and without the use of momentum, is consistent with standard theory in stochastic approximation and asymptotic convergence with probability one holds for the method under the most general setting of asynchrony.", "We presented numerical results that indicate some possible performance variabilities in three types of asynchrony: block partitioning inconsistent read (for which the above convergence theory applies), full-variable-update consistent write (for which the above convergence theory also applies), and full-variable-update inconsistent read/write (for which no convergence theory exists)." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.060606054961681366, 0.1304347813129425, 0, 0.17142856121063232, 0.1875, 0.11428570747375488, 0, 0.05882352590560913, 0, 0.17142856121063232, 0.0416666604578495, 0, 0.08510638028383255, 0.05405404791235924, 0.11764705181121826, 0.0363636314868927, 0.05882352590560913, 0.1599999964237213, 0.045454539358615875, 0, 0.1395348757505417, 0.31111109256744385, 0, 0.10526315122842789, 0, 0.08695651590824127, 0.07843136787414551, 0.17391303181648254, 0.12244897335767746, 0, 0.3243243098258972, 0.0624999962747097, 0.25641024112701416, 0.3243243098258972, 0.0952380895614624, 0.3333333432674408, 0.04999999701976776, 0.12765957415103912, 0.10526315122842789, 0.1304347813129425, 0.20689654350280762, 0.31578946113586426, 0.03999999538064003 ]

[ "Asymptotic convergence for stochastic subgradien method with momentum under general parallel asynchronous computation for general nonconvex nonsmooth optimization" ]

[ "Stochastic neural networks with discrete random variables are an important class of models for their expressivity and interpretability.", "Since direct differentiation and backpropagation is not possible, Monte Carlo gradient estimation techniques have been widely employed for training such models.", "Efficient stochastic gradient estimators, such Straight-Through and Gumbel-Softmax, work well for shallow models with one or two stochastic layers.", "Their performance, however, suffers with increasing model complexity.\n", "In this work we focus on stochastic networks with multiple layers of Boolean latent variables.", "To analyze such such networks, we employ the framework of harmonic analysis for Boolean functions. ", "We use it to derive an analytic formulation for the source of bias in the biased Straight-Through estimator.", "Based on the analysis we propose \\emph{FouST}, a simple gradient estimation algorithm that relies on three simple bias reduction steps.", "Extensive experiments show that FouST performs favorably compared to state-of-the-art biased estimators, while being much faster than unbiased ones.", "To the best of our knowledge FouST is the first gradient estimator to train up very deep stochastic neural networks, with up to 80 deterministic and 11 stochastic layers.", "Stochastic neural networks with discrete latent variables have been an alluring class of models for their expressivity and interpretability, dating back to foundational work on Helmholtz machines (Dayan et al., 1995) and sigmoid belief nets (Neal, 1992) .", "Since they are not directly differentiable, discrete random variables do not mesh well with the workhorse of modern Deep Learning, that is the backpropagation algorithm.", "Monte Carlo gradient estimation is an effective solution where, instead of computing the true gradients, one can sample gradients from some distribution.", "The sample estimates can be either biased or unbiased.", "Unbiased gradient estimates like score function estimators (Williams, 1992) come typically at the cost of high variance leading to slow learning.", "In contrast, biased gradient estimates such Straight-Through (Bengio et al., 2013) , while efficient, run the risk of convergence to poor minima and unstable training.", "To this end several solutions have recently been proposed that either reduce variance in unbiased estimators (Mnih & Gregor, 2014; Gu et al., 2015; Tucker et al., 2017; Rezende et al., 2014; Grathwohl et al., 2017) or control bias in biased estimators (Jang et al., 2016; Maddison et al., 2016) .", "These methods, however, have difficulty scaling up to complex neural networks with multiple stochastic layers: low-variance unbiased estimators are too expensive 1 , while the compounded bias from the continuous relaxations on multiple stochastic layers leads to poor minima.", "In this work we focus on biased estimators.", "Our goal in this paper is a gradient estimator for Boolean random variables that works for any complex -deep or wide-neural network architecture.", "We resort to the term Boolean instead of binary to emphasize that we work directly on the Boolean space {−1, +1}, without any continuous relaxations or quantizations.", "With this in mind we re-purpose the framework of harmonic analysis of Boolean functions, widely used in computational learning and computational complexity theory (O'Donnell, 2014; Linial et al., 1993; Mossel et al., 2003; Mansour, 1994) .", "We cast stochastic neural networks as Boolean functions f (z) over Boolean latent variables z sampled from probability 1.", "We introduce the framework of harmonic analysis of Boolean functions to analyze discrete stochastic neural networks and their REINFORCE and Straight-Through gradients.", "We show that stochastic gradients compute Fourier coefficients.", "2. Based on the above harmonic analysis we present FouST -a low-bias gradient estimator for Boolean latent variables based on three bias reduction steps.", "As a side contribution, we show that the gradient estimator employed with DARN (Gregor et al., 2013) , originally proposed for autoregressive models, is a strong baseline for gradient estimation in large and complex models with many stochastic layers.", "3. We show that FouST is amenable to complex stochastic neural networks with Boolean random variables.", "To the best of our knowledge, FouST is the first gradient estimate algorithm that can train very deep stochastic neural networks with Boolean latent variables.", "The practical outcome is a simple gradient estimate algorithm that can be plugged in complex stochastic neural networks with multiple layers of Boolean random variables." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0 ]

[ 0.19354838132858276, 0.11764705181121826, 0.32258063554763794, 0.09090908616781235, 0.2857142686843872, 0.1428571343421936, 0.19999998807907104, 0.06451612710952759, 0, 0.21052631735801697, 0.11999999731779099, 0.0555555522441864, 0, 0, 0, 0, 0, 0.1249999925494194, 0, 0.22857142984867096, 0.10810810327529907, 0.045454543083906174, 0.19354838132858276, 0.1818181723356247, 0.1904761791229248, 0.277777761220932, 0.3333333432674408, 0.27586206793785095, 0.1621621549129486, 0.2631579041481018 ]

[ "We present a low-bias estimator for Boolean stochastic variable models with many stochastic layers." ]

[ "We propose a novel generative adversarial network for visual attributes manipulation (ManiGAN), which is able to semantically modify the visual attributes of given images using natural language descriptions.", "The key to our method is to design a novel co-attention module to combine text and image information rather than simply concatenating two features along the channel direction.", "Also, a detail correction module is proposed to rectify mismatched attributes of the synthetic image, and to reconstruct text-unrelated contents.", "Finally, we propose a new metric for evaluating manipulation results, in terms of both the generation of text-related attributes and the reconstruction of text-unrelated contents.", "Extensive experiments on benchmark datasets demonstrate the advantages of our proposed method, regarding the effectiveness of image manipulation and the capability of generating high-quality results.", "Image manipulation refers to the task of changing various aspects of given images from low-level colour or texture Gatys et al., 2016) to high-level semantics (Zhu et al., 2016) , and has numerous potential applications in video games, image editing, and computer-aided design.", "Recently, with the development of deep learning and generative models, automatic image manipulation becomes possible, including image inpainting (Iizuka et al., 2016; Pathak et al., 2016) , image colourisation , style transfer (Gatys et al., 2016; Johnson et al., 2016) , and domain or attribute translation (Lample et al., 2017; .", "However, all the above works mainly focus on specific tasks, and only few studies (Dong et al., 2017; Nam et al., 2018) concentrate on more general and user-friendly image manipulation by using natural language descriptions.", "Also, as shown in Fig.1 , current state-of-the-art methods can only generate low-quality images and fail to effectively manipulate given images on more complicated datasets, such as COCO (Lin et al., 2014) .", "The less effective performance is mainly because (1) simply concatenating text and image cross-domain features along the channel direction, the model may fail to precisely correlate words and corresponding visual attributes, and thus cannot modified specific attributes required in the text, and (2) conditioned only on a global sentence vector, current state-of-the-art methods lack important fine-grained information at the word-level, which prevents an effective manipulation using natural language descriptions.", "In this paper, we aim to manipulate given images using natural language descriptions.", "In particular, we focus on modifying visual attributes (e.g., category, texture, colour, and background) of input images by providing texts that describe desired attributes.", "To achieve this, we propose a novel generative adversarial network for visual attributes manipulation (ManiGAN), which allows to effectively manipulate given images using natural language descriptions and to produce high-quality results.", "The contribution of our proposed method is fourfold: (1) instead of simply concatenating hidden features generated from a natural language description and image features encoded from the input image along the channel direction, we propose a novel co-attention module where both features can collaborate to reconstruct the input image and also keep the synthetic result semantically aligned with the given text description, (2) a detail correction module (DCM) is introduced to rectify mismatched attributes, and to reconstruct text-unrelated contents existing in the input image, (3) a new metric is proposed, which can appropriately reflect the generation of text-related visual attributes and the reconstruction of text-unrelated contents involved in the image manipulation, and (4) extensive experiments on the CUB (Wah et al., 2011) and COCO (Lin et al., 2014) Figure 1: Examples of image manipulation using natural language descriptions.", "Current state-of-theart methods only generate low-quality images, and fail to do manipulation on COCO.", "In contrast, our method allows the input images to be manipulated accurately corresponding to the given text descriptions while preserving text-unrelated contents.", "to demonstrate the superiority of our model, which outperforms existing state-of-the-art methods both qualitatively and quantitatively.", "We have proposed a novel generative adversarial network for visual attributes manipulation, called ManiGAN, which can semantically manipulate the input images using natural language descriptions.", "Two novel components are proposed in our model: (1) the co-attention module enables cooperation between hidden features and image features where both features can collaborate to reconstruct the input image and also keep the synthetic result semantically aligned with the given text description, and (2) the detail correction module can rectify mismatched visual attributes of the synthetic result, and also reconstruct text-unrelated contents existing in the input image.", "Extensive experimental results demonstrate the superiority of our proposed method, in terms of both the effectiveness of image manipulation and the capability of generating high-quality results." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.5641025304794312, 0.20512820780277252, 0.1249999925494194, 0.11428570747375488, 0, 0.11999999731779099, 0, 0.17777776718139648, 0.1818181723356247, 0.1599999964237213, 0.6153846383094788, 0.052631575614213943, 0.5116279125213623, 0.18518517911434174, 0.07407406717538834, 0.3030303120613098, 0.06896550953388214, 0.4736841917037964, 0.09836065024137497, 0 ]

[ "We propose a novel method to manipulate given images using natural language descriptions." ]

[ "Optimal Transport (OT) naturally arises in many machine learning applications, where we need to handle cross-modality data from multiple sources.", "Yet the heavy computational burden limits its wide-spread uses.", "To address the scalability issue, we propose an implicit generative learning-based framework called SPOT (Scalable Push-forward of Optimal Transport).", "Specifically, we approximate the optimal transport plan by a pushforward of a reference distribution, and cast the optimal transport problem into a minimax problem.", "We then can solve OT problems efficiently using primal dual stochastic gradient-type algorithms.", "We also show that we can recover the density of the optimal transport plan using neural ordinary differential equations.", "Numerical experiments on both synthetic and real datasets illustrate that SPOT is robust and has favorable convergence behavior.", "SPOT also allows us to efficiently sample from the optimal transport plan, which benefits downstream applications such as domain adaptation.", "The Optimal Transport (OT) problem naturally arises in a variety of machine learning applications, where we need to handle cross-modality data from multiple sources.", "One example is domain adaptation: We collect multiple datasets from different domains, and we need to learn a model from a source dataset, which can be further adapted to target datasets BID18 BID8 .", "Another example is resource allocation: We want to assign a set of assets (one data source) to a set of receivers (another data source) so that an optimal economic benefit is achieved BID46 BID17 .", "Recent literature has shown that both aforementioned applications can be formulated as optimal transport problems.The optimal transport problem has a long history, and its earliest literature dates back to Monge (1781).", "Since then, it has attracted increasing attention and been widely studied in multiple communities such as applied mathematics, probability, economy and geography BID51 Carlier, 2012; BID23 .", "Specifically, we consider two sets of data, which are generated from two different distributions denoted by X ∼ µ and Y ∼ ν.1", "We aim to find an optimal joint distribution γ of X and Y , which minimizes the expectation on some ground cost function c, i.e., γ * = arg min γ∈Π(µ,ν) DISPLAYFORM0 The constraint γ ∈ Π(µ, ν) requires the marginal distribution of X and Y in γ to be identical to µ and ν, respectively.", "The cost function c measures the discrepancy between input X and Y .", "For crossmodality structured data, the form of c incorporates prior knowledge into optimal transport problem.", "Existing literature often refers to the optimal expected cost W * (µ, ν) = E (X,Y )∼γ * [c(X, Y )] as Wasserstein distance when c is a distance, and γ * as the optimal transport plan.", "For domain adaptation, the function c measures the discrepancy between X and Y , and the optimal transport plan γ * essentially reveals the transfer of the knowledge from source X to target Y .", "For resource allocation, the function c is the cost of assigning resource X to receiver Y , and the optimal transport plan γ To address the scalability and efficiency issues, we propose a new implicit generative learning-based framework for solving optimal transport problems.", "Specifically, we approximate γ * by a generative model, which maps from some latent variable Z to (X, Y ).", "For simplicity, we denote DISPLAYFORM1 where ρ is some simple latent distribution and G is some operator, e.g., deep neural network or neural ordinary differential equation (ODE) .", "Accordingly, instead of directly estimating the probability density of γ * , we estimate the mapping G between Z and (X, Y ) by solving DISPLAYFORM2 We then cast equation 3 into a minimax optimization problem using the Lagrangian multiplier method.", "As the constraints in equation 3 are over the space of continuous distributions, the Lagrangian multiplier is actually infinite dimensional.", "Thus, we propose to approximate the Lagrangian multiplier by deep neural networks, which eventually delivers a finite dimensional generative learning problem.Our proposed framework has three major benefits: (1) Our formulated minimax optimization problem can be efficiently solved by primal dual stochastic gradient-type algorithms.", "Many empirical studies have corroborated that these algorithms can easily scale to very large minimax problems in machine learning BID2 ; (2) Our framework can take advantage of recent advances in deep learning.", "Many empirical evidences have suggested that deep neural networks can effectively adapt to data with intrinsic low dimensional structures BID33 .", "Although they are often overparameterized, due to the inductive biases of the training algorithms, the intrinsic dimensions of deep neural networks are usually controlled very well, which avoids the curse of dimensionality; (3) Our adopted generative models allow us to efficiently sample from the optimal transport plan.", "This is very convenient for certain downstream applications such as domain adaptation, where we can generate infinitely many data points paired across domains BID35 .Moreover", ", the proposed framework can also recover the density of entropy regularized optimal transport plan. Specifically", ", we adopt the neural Ordinary Differential Equation (ODE) approach in to model the dynamics that how Z gradually evolves to G(Z). We then derive", "the ODE that describes how the density evolves, and solve the density of the transport plan from the ODE. The recovery of", "density requires no extra parameters, and can be evaluated efficiently.Notations: Given a matrix A ∈ R d×d , det(A) denotes its determinant, tr(A) = i A ii denotes its trace, A F = i,j A 2 ij denotes its Frobenius norm, and |A| denotes a matrix with [|A|] ij = |A ij |. We use dim(v) to", "denote the dimension of a vector v.", "Existing literature shows that several stochastic algorithms can efficiently compute the Wasserstein distance between two continuous distributions.", "These algorithms, however, only apply to the dual of the OT problem equation 1, and cannot provide the optimal transport plan.", "For example, BID19 suggest to expand the dual variables in two reproducing kernel Hilbert spaces.", "They then apply the Stochastic Averaged Gradient (SAG) algorithm to compute the optimal objective value of OT with continuous marginal distributions or semi-discrete marginal distributions (i.e., one marginal distribution is continuous and the other is discrete).", "The follow-up work, BID47 , parameterize the dual variables with neural networks and apply the Stochastic Gradient Descent (SGD) algorithm to eventually achieve a better convergence.", "These two methods can only provide the optimal transport plan and recover the joint density when the densities of the marginal distributions are known.", "This is prohibitive in most applications, since we only have access to the empirical data.", "Our framework actually allows us to efficiently compute the joint density from the transformation of the latent variable Z as in Section 4.", "TAB1 shows the architecture of two discriminators λ X , λ Y .", "The two networks have identical architechture and do not share parameters.", "The CNN architecture for USPS, MNIST and MNISTM.", "PReLU activation is applied BID24 .", "TAB2 shows the architecture of two generators G X and G Y .", "The last column in TAB2 means whether G X and G Y share the same parameter.", "TAB3 shows the architecture of two discriminators λ X , λ Y , and two classifiers D X , D Y .", "The last column in TAB2 uses (·, ·) to denote which group of discriminators share the same parameter.", "TAB4 shows the architecture of two generators G X and G Y .", "The last column in TAB4 means whether G X and G Y share the same parameter.", "The Residual block is the same as the one in BID38 .", "[3 × 3, ch, stride = 1, padding =0] Sigmoid False TAB5 shows the architecture of two discriminators λ X , λ Y , and two classifiers D X , D Y .", "The last column in TAB5 uses (·, ·) to denote which group of discriminators share the same parameter." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.0714285671710968, 0, 0, 0.1538461446762085, 0.0952380895614624, 0.1538461446762085, 0, 0.2142857164144516, 0.0624999962747097, 0.054054051637649536, 0.11428570747375488, 0.1666666567325592, 0, 0, 0.07547169923782349, 0, 0.17391303181648254, 0.14999999105930328, 0.17142856121063232, 0.13636364042758942, 0.0714285671710968, 0, 0.043478257954120636, 0, 0.04081632196903229, 0.052631575614213943, 0.0714285671710968, 0.12765957415103912, 0, 0.17391303181648254, 0.06451612710952759, 0.17391303181648254, 0.039215683937072754, 0, 0, 0.2222222238779068, 0.08695651590824127, 0.10256409645080566, 0.060606058686971664, 0.13793103396892548, 0.08695651590824127, 0.06896550953388214, 0, 0, 0, 0, 0, 0, 0, 0.07692307233810425, 0, 0, 0, 0, 0.07692307233810425 ]

[ "Use GAN-based method to scalably solve optimal transport" ]

[ "In this work, we propose a novel formulation of planning which views it as a probabilistic inference problem over future optimal trajectories.", "This enables us to use sampling methods, and thus, tackle planning in continuous domains using a fixed computational budget. ", "We design a new algorithm, Sequential Monte Carlo Planning, by leveraging classical methods in Sequential Monte Carlo and Bayesian smoothing in the context of control as inference.", "Furthermore, we show that Sequential Monte Carlo Planning can capture multimodal policies and can quickly learn continuous control tasks.", "To exhibit intelligent behaviour machine learning agents must be able to learn quickly, predict the consequences of their actions, and explain how they will react in a given situation.", "These abilities are best achieved when the agent efficiently uses a model of the world to plan future actions.", "To date, planning algorithms have yielded very impressive results.", "For instance, Alpha Go BID36 relied on Monte Carlo Tree Search (MCTS) BID23 ) to achieve super human performances.", "Cross entropy methods (CEM) BID34 have enabled robots to perform complex nonprehensile manipulations BID11 and algorithms to play successfully Tetris BID39 .", "In addition, iterative linear quadratic regulator (iLQR) BID21 BID20 BID41 enabled humanoid robots tasks to get up from an arbitrary seated pose .Despite", "these successes, these algorithms make strong underlying assumptions about the environment. First,", "MCTS requires a discrete setting, limiting most of its successes to discrete games with known dynamics. Second", ", CEM assumes the distribution over future trajectories to be Gaussian, i.e. unimodal. Third", ", iLQR assumes that the dynamics are locally linear-Gaussian, which is a strong assumption on the dynamics and would also assume the distribution over future optimal trajectories to be Gaussian. For", "these reasons, planning remains an open problem in environments with continuous actions and complex dynamics. In", "this paper, we address the limitations of the aforementioned planning algorithms by creating a more general view of planning that can leverage advances in deep learning (DL) and probabilistic inference methods. This", "allows us to approximate arbitrary complicated distributions over trajectories with non-linear dynamics.We frame planning as density estimation problem over optimal future trajectories in the context of control as inference BID10 BID45 BID43 Rawlik et al., 2010; BID47 BID31 . This", "perspective allows us to make use of tools from the inference research community and, as previously mentioned, model any distribution over future trajectories. The", "planning distribution is complex since trajectories consist of an intertwined sequence of states and actions. Sequential", "Monte Carlo (SMC) BID38 BID13 BID27 methods are flexible and efficient to model such a T −1 t≥1 p env (s t+1 |s t , a t ) T t≥1 π θ (a t |s t ) denotes the probability of a trajectory x 1:T under policy π θ . FIG8 .1: O", "t is an observed optimality variable with probability p(O t |s t , a t ) = exp(r(s t , a t )).x t = (s t", ", a t ) are the state-action pair variables considered here as latent.Traditionally, in reinforcement learning (RL) problems, the goal is to find the optimal policy that maximizes the expected return E q θ [ T t=1 γ t r t ]. However, it", "is useful to frame RL as an inference problem within a probabilistic graphical framework BID33 BID45 BID30 . First, we introduce", "an auxiliary binary random variable O t denoting the \"optimality\" of a pair (s t , a t ) at time t and define its probability 1 as p(O t = 1|s t , a t ) = exp(r(s t , a t )). O is a convenience", "variable only here for the sake of modeling. By considering the", "variables (s t , a t ) as latent and O t as observed, we can construct a Hidden Markov Model (HMM) as depicted in figure 2.1. Notice that the link", "s → a is not present in figure 2.1 as the dependency of the optimal action on the state depends on the future observations. In this graphical model", ", the optimal policy is expressed as p(a t |s t , O t:T ).The posterior probability", "of this graphical model can be written as 2 : DISPLAYFORM0 r(s t , a t ) + log p(a t ) .(2.1)It appears clearly that", "finding optimal trajectories is equivalent to finding plausible trajectories yielding a high return.1 as in BID30 , if the rewards are bounded above, we can always remove a constant so that the probability is well defined.2 Notice that in the rest of the", "paper, we will abusively remove the product of the action priors T t=1 p(at) = exp T t=1 log p(at) from the joint as in BID30 . We typically consider this term", "either constant or already included in the reward function. See Appendix A.2 for details.Many", "control as inference methods can be seen as approximating the density by optimizing its variational lower bound: BID43 . Instead of directly differentiating", "the variational lower bound for the whole trajectory, it is possible to take a message passing approach such as the one used in Soft Actor-Critic (SAC) BID17 and directly estimate the optimal policy p(a t |s t , O t:T ) using the backward message, i.e a soft Q function instead of the Monte Carlo return. DISPLAYFORM1", "In this work, we have introduced a connection between planning and inference and showed how we can exploit advances in deep learning and probabilistic inference to design a new efficient and theoretically grounded planning algorithm.", "We additionally proposed a natural way to combine model-free and model-based reinforcement learning for planning based on the SMC perspective.", "We empirically demonstrated that our method achieves state of the art results on Mujoco.", "Our result suggest that planning can lead to faster learning in control tasks.However, our particle-based inference method suffers some several shortcomings.", "First, we need many particles to build a good approximation of the posterior, and this can be computationally expensive since it requires to perform a forward pass of the policy, the value function and the model for every particle.", "Second, resampling can also have adverse effects, for instance all the particles could be resampled on the most likely particle, leading to a particle degeneracy.", "More advanced SMC methods dealing with this issue such as backward simulation BID32 or Particle Gibbs with Ancestor Sampling (PGAS) (Lindsten et al., 2014) have been proposed and using them would certainly improve our results.Another issue we did not tackle in our work is the use of models of the environment learned from data.", "Imperfect model are known to result in compounding errors for prediction over long sequences.", "We chose to re-plan at each time step (Model Predictive Control) as it is often done in control to be more robust to model errors.", "More powerful models or uncertainty modeling techniques can also be used to improve the accuracy of our planning algorithm.", "While the inference and modeling techniques used here could be improved in multiple ways, SMCP achieved impressive learning speed on complex control tasks.", "The planning as inference framework proposed in this work is general and could serve as a stepping stone for further work combining probabilistic inference and deep reinforcement learning." ]

[ 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.3333333432674408, 0.22857142984867096, 0.42105263471603394, 0.3636363446712494, 0.09090908616781235, 0.060606054961681366, 0.0833333283662796, 0.11764705181121826, 0.05714285373687744, 0, 0, 0.06451612710952759, 0, 0.09302324801683426, 0.12903225421905518, 0.27272728085517883, 0.15094339847564697, 0.10256409645080566, 0.19999998807907104, 0.1428571343421936, 0.060606054961681366, 0.072727270424366, 0.2857142686843872, 0.1304347813129425, 0, 0.1904761791229248, 0.09999999403953552, 0.06451612710952759, 0.10256409645080566, 0.11764705181121826, 0.09756097197532654, 0, 0.1666666567325592, 0.1515151411294937, 0.3255814015865326, 0.22857142984867096, 0, 0.1621621549129486, 0.12765957415103912, 0.05128204822540283, 0.12121211737394333, 0, 0.10526315122842789, 0.11764705181121826, 0.15789473056793213, 0.3589743673801422 ]

[ "Leveraging control as inference and Sequential Monte Carlo methods, we proposed a probabilistic planning algorithm." ]

[ "Training Generative Adversarial Networks (GANs) is notoriously challenging.", "We propose and study an architectural modification, self-modulation, which improves GAN performance across different data sets, architectures, losses, regularizers, and hyperparameter settings.", "Intuitively, self-modulation allows the intermediate feature maps of a generator to change as a function of the input noise vector.", "While reminiscent of other conditioning techniques, it requires no labeled data.", "In a large-scale empirical study we observe a relative decrease of 5%-35% in FID.", "Furthermore, all else being equal, adding this modification to the generator leads to improved performance in 124/144 (86%) of the studied settings.", "Self-modulation is a simple architectural change that requires no additional parameter tuning, which suggests that it can be applied readily to any GAN.", "Generative Adversarial Networks (GANs) are a powerful class of generative models successfully applied to a variety of tasks such as image generation BID20 Miyato et al., 2018; Karras et al., 2017) , learned compression BID15 , super-resolution (Ledig et al., 2017) , inpainting (Pathak et al., 2016) , and domain transfer BID13 BID23 .Training", "GANs is a notoriously challenging task BID6 BID15 as one is searching in a high-dimensional parameter space for a Nash equilibrium of a non-convex game. As a practical", "remedy one applies (usually a variant of) stochastic gradient descent, which can be unstable and lack guarantees Salimans et al. (2016) . As a result, one", "of the main research challenges is to stabilize GAN training. Several approaches", "have been proposed, including varying the underlying divergence between the model and data distributions Mao et al., 2016) , regularization and normalization schemes BID7 Miyato et al., 2018) , optimization schedules (Karras et al., 2017) , and specific neural architectures (Radford et al., 2016; BID21 . A particularly successful", "approach is based on conditional generation; where the generator (and possibly discriminator) are given side information, for example class labels Mirza & Osindero (2014) ; Odena et al. (2017) ; Miyato & Koyama (2018) . In fact, state-of-the-art", "conditional GANs inject side information via conditional batch normalization (CBN) layers BID3 Miyato & Koyama, 2018; BID21 . While this approach does", "help, a major drawback is that it requires external information, such as labels or embeddings, which is not always available.In this work we show that GANs benefit from self-modulation layers in the generator. Our approach is motivated", "by Feature-wise Linear Modulation in supervised learning (Perez et al., 2018; BID3 , with one key difference: instead of conditioning on external information, we condition on the generator's own input. As self-modulation requires", "a simple change which is easily applicable to all popular generator architectures, we believe that is a useful addition to the GAN toolbox.", "We present a generator modification that improves the performance of most GANs.", "This technique is simple to implement and can be applied to all popular GANs, therefore we believe that selfmodulation is a useful addition to the GAN toolbox.Our results suggest that self-modulation clearly yields performance gains, however, they do not say how this technique results in better models.", "Interpretation of deep networks is a complex topic, especially for GANs, where the training process is less well understood.", "Rather than purely speculate, we compute two diagnostic statistics that were proposed recently ignite the discussion of the method's effects.First, we compute the condition number of the generators Jacobian.", "Odena et al. (2018) provide evidence that better generators have a Jacobian with lower condition number and hence regularize using this quantity.", "We estimate the generator condition number in the same was as Odena et al. (2018) .", "We compute the Jacobian (J z ) i,j = δG(z)i δzj at each z in a minibatch, then average the logarithm of the condition numbers computed from each Jacobian.Second, we compute a notion of precision and recall for generative models.", "Sajjadi et al. (2018) define the quantities, F 8 and F 1/8 , for generators.", "These quantities relate intuitively to the traditional precision and recall metrics for classification.", "Generating points which have low probability under the true data distribution is interpreted as a loss in precision, and is penalized by the F 8 score.", "Failing to generate points that have high probability under the true data distributions is interpreted as a loss in recall, and is penalized by the F 1/8 score.", "FIG4 shows both statistics.", "The left hand plot shows the condition number plotted against FID score for each model.", "We observe that poor models tend to have large condition numbers; the correlation, although noisy, is always positive.", "This result corroborates the observations in (Odena et al., 2018) .", "However, we notice an inverse trend in the vicinity of the best models.", "The cluster of the best models with self-modulation has lower FID, but higher condition number, than the best models without self-modulation.", "Overall the correlation between FID and condition number is smaller for self-modulated models.", "This is surprising, it appears that rather than unilaterally reducing the condition number, self-modulation provides some training stability, yielding models with a small range of generator condition numbers.The right-hand plot in FIG4 shows the F 8 and F 1/8 scores.", "Models in the upper-left quadrant cover true data modes better (higher precision), and models in the lower-right quadrant produce more modes (higher recall).", "Self-modulated models tend to favor higher recall.", "This effect is most pronounced on IMAGENET.", "Overall these diagnostics indicate that self-modulation stabilizes the generator towards favorable conditioning values.", "It also appears to improve mode coverage.", "However, these metrics are very new; further development of analysis tools and theoretical study is needed to better disentangle the symptoms and causes of the self-modulation technique, and indeed of others.Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen.", "Progressive growing of gans for improved quality, stability, and variation.", "A ADDITIONAL RESULTS" ]

[ 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0, 0.37837836146354675, 0, 0, 0, 0.1111111044883728, 0.15789473056793213, 0.03448275476694107, 0, 0.05128204822540283, 0.0714285671710968, 0.11320754140615463, 0, 0, 0.039215680211782455, 0, 0.2222222238779068, 0.2857142686843872, 0.17241379618644714, 0, 0.04999999701976776, 0.10526315122842789, 0, 0.04081632196903229, 0.06666666269302368, 0.06896550953388214, 0.04999999701976776, 0.0952380895614624, 0, 0, 0.05882352590560913, 0, 0, 0, 0.06896550953388214, 0.07407406717538834, 0.05882352590560913, 0, 0, 0.06896550953388214, 0, 0.03999999538064003, 0.07692307233810425, 0.10526315867900848 ]

[ "A simple GAN modification that improves performance across many losses, architectures, regularization schemes, and datasets. " ]

[ "Extreme Classification Methods have become of paramount importance, particularly for Information Retrieval (IR) problems, owing to the development of smart algorithms that are scalable to industry challenges.", "One of the prime class of models that aim to solve the memory and speed challenge of extreme multi-label learning is Group Testing.", "Multi-label Group Testing (MLGT) methods construct label groups by grouping original labels either randomly or based on some similarity and then train smaller classifiers to first predict the groups and then recover the original label vectors.", "Recently, a novel approach called MACH (Merged Average Classifiers via Hashing) was proposed which projects the huge label vectors to a small and manageable count-min sketch (CMS) matrix and then learns to predict this matrix to recover the original prediction probabilities.", "Thereby, the model memory scales O(logK) for K classes.", "MACH is a simple algorithm which works exceptionally well in practice.", "Despite this simplicity of MACH, there is a big gap between the theoretical understanding of the trade-offs with MACH.", "In this paper we fill this gap.", "Leveraging the theory of count-min sketch we provide precise quantification of the memory-identifiablity tradeoffs.", "We extend the theory to the case of multi-label classification, where the dependencies make the estimators hard to calculate in closed forms.", "To mitigate this issue, we propose novel quadratic approximation using the Inclusion-Exclusion Principle.", "Our estimator has significantly lower reconstruction error than the typical CMS estimator across various values of number of classes K, label sparsity and compression ratio.", "Extreme Classification has taken center-stage of Data Mining and Information Retrieval research in the past few years (Zamani et al., 2018; Prabhu et al., 2018b; Jain et al., 2019; Choromanska & Langford, 2015) .", "It refers to the vanilla multiclass and multilabel classification problems where the number of classes K is significantly large.", "A large number of classes K brings a new set of computational and memory challenges in training and deploying classifiers.", "There have been several paradigms of models that tackle the scale challenge of Extreme Classification like 1-vs-all methods (Prabhu et al., 2018b; Jain et al., 2019; Babbar & Schölkopf, 2017) , tree based methods (Prabhu et al., 2018a; Jain et al., 2016) , embedding models (Nigam et al., 2019; Bhatia et al., 2015) , etc. (as noted on the popular Extreme Classification Repository).", "One of the recent approaches proposed to alleviate the scale challenge of Multilabel Classification is Group Testing (Ubaru & Mazumdar, 2017; Ubaru et al., 2016; Vem et al., 2017) .", "In this method, all labels are grouped randomly into m groups/clusters.", "Each label may go into more than one group.", "We first train a classifier that predicts which of these clusters the input belongs to (treating each cluster as a separate label in a multilabel setting).", "For any given input, we first predict the clusters into which the true labels of the input may have been pooled.", "We can then identify all the true labels by taking an intersection over the inverted clusters.", "This approach suffers from a critical problem that even tree based approaches have, i.e., hard assignment of clusters.", "Since the recovery of true labels depends solely on hard-prediction of clusters, a mistake in the cluster prediction can cost us dearly in the final label prediction.", "Also, since the labels are pooled randomly, each individual meta-classifier is a weak and noisy one.", "In a recent development, Merged Average Classifiers via Hashing (MACH) (Medini et al., 2019) was proposed that alleviates the hard-prediction problem in Group Testing methods by identifying the best labels based on the sum of prediction probabilities of the respective groups for a given input.", "In the hindsight, MACH subtly learns to predict a count-min sketch (CMS) (Cormode & Muthukrishnan, 2005 ) matrix of the original probability vector.", "For the case of multiclass classification (every input having just a single label unlike multilabel), MACH proposes an unbiased estimator to recover the original K dimensional probability vector from the predicted CMS matrix.", "Multiclass classification naturally fits into the count-min sketch setting as no two labels can appear simultaneously for a given input.", "But the proposed theory does not naturally extend to multilabel learning.", "Further, the variance and error bounds for multiclass classification rely heavily on the choice of number of hash tables and the size of each hash table.", "That aspect has not been explored in prior work.", "Our Contributions: In this work we broadly make the following contributions:", "1) We revisit MACH with a thorough analysis of proposed reconstruction estimator for multiclass learning.", "In particular, we prove that the variance of estimation is inversely proportional to the product of product of number of hash tables and size of each hash table (in theorem 2).", "2) We also obtain a lower bound on hash table hyperparametrs given a tolerance to prediction error (in Theorems 4 and 5).", "3) We propose a novel reconstruction estimator for the case of multilabel learning using InclusionExclusion principle (in theorem 6).", "This estimator comes out as a solution to a quadratic equation (hence we code-name it as 'quadratic estimator').", "4) We simulate multilabel learning setting by generating K dimensional probability vectors and their proxy CMS measurements.", "We then reconstruct the probability vector using both the mean estimator and the quadratic estimator and show that the reconstruction Mean-Squared Error (MSE) is significantly lower for the new estimator.", "Following the above steps, we show the comparison of our proposed quadratic estimator in theorem 6 against the plain mean estimator by varying the values of K, B, V and base prob in figure 3 .", "We can infer the following insights from the plots :", "• As K increases, the MSE grows.", "This is expected because the reconstructed vector has a small non-zero probability for many of the K classes and this induces noise and hence MSE grows.", "But the top classes are still retrieved with high certainty.", "• For any K, V, base prob, the MSE decreases when B increases which is expected (fewer collisions of classes and hence less noisier predictions).", "As the MSE gets lower, the gains from the square-root estimator are also low.", "This is good because in scenarios where B and R are small, we can do much better recovery using the proposed estimator.", "• For any K, B, base prob the MSE increases with V .", "This is again natural because larger V induces more 'true' class collisions and hence the retrieval becomes fuzzy.", "• For any K, B, V the MSE decreases with base prob, albeit with much little difference than previous cases.", "This is interesting because a high base prob means that we have few but highly confident 'true' classes among K. On the other hand, lower base prob indicates that 'true' classes are scattered among a larger subset among K classes.", "Yet, MACH recovers the original probabilities with commendably low MSE.", "Varying B for K = 10000", "Varying base prob for K = 10000", "Varying B for K = 100000 Varying V for K = 100000 Varying base prob for K = 100000", "Varying B for K = 1000000 Varying B for K = 1000000 Varying base prob for K = 1000000", "Figure 3: Reconstruction Error (MSE) comparison between", "1) vanilla mean estimator (plotted in magenta) and", "2) proposed square-root estimator (plotted in green); for various configurations of K,B and V. The value of K varies as 10000, 100000, 1000000 for the 1 st , 2 nd and 3 rd rows respectively.", "In each row, the first plot fixes V, base prob and compares various values of B. The 2 nd plot fixes B, base prob and compares different values of B. The 3 rd one fixes B, V and compares different values of base prob.", "In all cases, we notice that the square-root estimator is consistently and significantly lower in MSE than the corresponding mean estimator.", "We perform a rigorous theoretical analysis of using Count-Min-Sketch for Extreme Classification and come up with error bounds and hyper-parameter constraints.", "We identify a critical shortcoming of reconstruction estimators proposed in prior research.", "We overcome the shortcoming by treating each bucket in a hash table as a union of merged original classes.", "Using inclusion-exclusion principle and a controlled label sparsity assumption, we come up with an approximate estimator to reconstruct original probability vector from the predicted Count-Min Sketch measurements.", "Our new estimator has significantly lower reconstruction MSE than the prior estimator." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.13333332538604736, 0.14999999105930328, 0.11999999731779099, 0.2181818187236786, 0.13793103396892548, 0.06451612710952759, 0.1621621549129486, 0, 0.1875, 0.10526315122842789, 0, 0.1395348757505417, 0.07999999821186066, 0.21052631735801697, 0.21052631735801697, 0.032258059829473495, 0.08695651590824127, 0, 0, 0.13636362552642822, 0.051282044500112534, 0, 0.14999999105930328, 0.0952380895614624, 0.1111111044883728, 0.09836065024137497, 0.380952388048172, 0.2745097875595093, 0.19999998807907104, 0.06451612710952759, 0.14999999105930328, 0, 0, 0.17142856121063232, 0.13636362552642822, 0.1463414579629898, 0.1538461446762085, 0.1111111044883728, 0.1621621549129486, 0.1860465109348297, 0.08163265138864517, 0.06896550953388214, 0, 0.3181818127632141, 0.06666666269302368, 0.13333332538604736, 0.0624999962747097, 0.0476190410554409, 0, 0.052631575614213943, 0, 0.07999999821186066, 0.06666666269302368, 0.07692307233810425, 0.07407406717538834, 0.06896550953388214, 0.0714285671710968, 0, 0.0714285671710968, 0.11538460850715637, 0.08888888359069824, 0.051282044500112534, 0.25, 0.1249999925494194, 0.21052631735801697, 0.3404255211353302, 0 ]

[ "How to estimate original probability vector for millions of classes from count-min sketch measurements - a theoretical and practical setup." ]

[ "Neural networks are commonly used as models for classification for a wide variety of tasks.", "Typically, a learned affine transformation is placed at the end of such models, yielding a per-class value used for classification.", "This classifier can have a vast number of parameters, which grows linearly with the number of possible classes, thus requiring increasingly more resources.\n\n", "In this work we argue that this classifier can be fixed, up to a global scale constant, with little or no loss of accuracy for most tasks, allowing memory and computational benefits.", "Moreover, we show that by initializing the classifier with a Hadamard matrix we can speed up inference as well.", "We discuss the implications for current understanding of neural network models.\n", "Deep neural network have become a widely used model for machine learning, achieving state-ofthe-art results on many tasks.", "The most common task these models are used for is to perform classification, as in the case of convolutional neural networks (CNNs) used to classify images to a semantic category.", "CNN models are currently considered the standard for visual tasks, allowing far better accuracy than preceding approaches BID21 BID42 .Training", "NN models and using them for inference requires large amounts of memory and computational resources, thus, extensive amount of research has been done lately to reduce the size of networks. BID6 used", "weight sharing and specification, BID28 used mixed precision to reduce the size of the neural networks by half. BID44 and", "BID19 used low rank approximations to speed up NNs. BID17 , BID24", "and BID51 , used a more aggressive approach, in which weights, activations and gradients were quantized to further reduce computation during training. Although aggressive", "quantization benefits from smaller model size, the extreme compression rate comes with a loss of accuracy.Past work noted the fact that predefined BID31 and random BID15 projections can be used together with a learned affine transformation to achieve competitive results on several tasks. In this study suggest", "the reversed proposal -that common NN models used can learn useful representation even without modifying the final output layer, which often holds a large number of parameters that grows linearly with number of classes.", "In this work we suggested removing the parameters from the classification layer used in deep neural networks.", "We showed empirical results suggesting that keeping the classifier fixed cause little or no decline in classification performance for common balanced datasets such as Cifar and Imagenet, while allowing a noticeable reduction in trainable parameters.", "We argue that fixing the last layer can reduce the computational complexity for training as well as the communication cost in distributed learning.", "Furthermore, using a Hadamard matrix as classifier might lead to some computational benefits when properly implemented, and save memory otherwise spent on large amount of transformation coefficients.", "As datasets tend to become more complex by time (e.g., Cifar100, ImageNet1K, ImageNet22k, JFT-300M, and language modeling) we believe that resource hungry affine transformation should remain fixed during training, at least partially.We also found that new efficient methods to create pre-defined word embeddings should be explored, as they require huge amount of parameters that can possibly be avoided when learning a new task.", "Based on these findings, we recommend future research to focus on representations learned by the non-linear part of neural networks -up to the final classifier, as it seems to be highly redundant." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0 ]

[ 0.07999999821186066, 0.06666666269302368, 0.1818181723356247, 0.1428571343421936, 0.20689654350280762, 0.17391303181648254, 0.06896550953388214, 0.21052631735801697, 0.12903225421905518, 0.09999999403953552, 0.20689654350280762, 0, 0.060606054961681366, 0.1071428507566452, 0.1428571343421936, 0.29629629850387573, 0.13333332538604736, 0.19354838132858276, 0.052631575614213943, 0.02816901169717312, 0.1538461446762085 ]

[ "You can fix the classifier in neural networks without losing accuracy" ]

[ "This paper introduces NEMO, an approach to unsupervised object detection that uses motion---instead of image labels---as a cue to learn object detection.", "To discriminate between motion of the target object and other changes in the image, it relies on negative examples that show the scene without the object.", "The required data can be collected very easily by recording two short videos, a positive one showing the object in motion and a negative one showing the scene without the object.", "Without any additional form of pretraining or supervision and despite of occlusions, distractions, camera motion, and adverse lighting, those videos are sufficient to learn object detectors that can be applied to new videos and even generalize to unseen scenes and camera angles.", "In a baseline comparison, unsupervised object detection outperforms off-the shelf template matching and tracking approaches that are given an initial bounding box of the object.", "The learned object representations are also shown to be accurate enough to capture the relevant information from manipulation task demonstrations, which makes them applicable to learning from demonstration in robotics.", "An example of object detection that was learned from 3 minutes of video can be found here: http://y2u.be/u_jyz9_ETz4", "Object-based representations are a powerful abstraction of our world.", "Since these representations remove large amounts of information-an image of size 120 × 160 for example has 120 × 160 × 3 = 57.600 dimensions, while the coordinates of an object in that image only have 2 dimensions-object-based representations enable efficient generalization, simulation, planning, communication, etc.", "But grounding objects in sensory input currently relies on supervised learning, which requires a high number of labeled images, e.g. 500.000 manually annotated segments to learn 80 objects BID22 .", "This paper takes a step towards replacing this labor-intensive supervision by learning to detect objects from videos that can be gathered quickly with minimal supervision and by exploiting the physical properties of objects.A physical object is a collection of matter that moves as a unit.", "Motion, in turn, can be a strong cue to learn object detection and replace the need for supervision in the form of labeled images.", "Given a video of a moving object, we can learn object-based representations by optimizing them to describe physically plausible motion BID16 .", "But this approach only works in the absence of visual distractions.", "With camera motion, other moving objects, or changes in the background, motion alone is not sufficient to learn such representations because many features in the image move in a physically plausible way.", "This paper improves on previous approaches by learning to ignore visual distractions through negative examples, i.e., videos of the scene without the target object but with the distractions.", "These negative videos are easy to collect because they do not need to be in sync with the positive ones, i.e., they do not need to have the same sequence of camera movements or the same object motions.", "This paper also addresses the challenge Figure 1 : Learning to detect an object from 3 min of video.", "Left to right: training video of a pen in hand, negative video without pen, two test videos with per frame detections shown as black dots.", "[video link] of changes between training and test videos, e.g. due to different lighting or changes in the background.", "Those changes can be harmful if an object representation is extracted using a standard pyramid-shaped convolutional network because every pixel directly affects the output, even if it is far from the object's location.", "Therefore, this paper uses a spatial encoder architecture that uses a spatial softmax output BID5 , which is only affected by the strongest local activations, making it invariant to many visual distractions.The contribution of this paper is to demonstrate unsupervised object detection based on data that is easy and fast to collect.", "This is achieved by formulating the use of negative examples for object detection as a loss function, combining it with motion-based learning objectives, and using these objectives to train a spatial encoder network using a combination of random search and gradient descent.", "The resulting method is called learning from negative examples and motion (NEMO).", "A glimpse of the results are shown in Figure 1 .Experimental", "results in Section 4 show that NEMO can learn new objects from only two short videos of a few minutes, without using pretrained models and without using supervision beyond marking these videos as positive and negative. The results", "also show that NEMO can learn object detection in the presence of frequent occlusions, distractions, camera motion, and changes in lighting and background. Even though", "it uses data that can be collected in seconds to minutes, the learned object detection generalizes to new scenes and camera angles and outperforms template matching and tracking baselines. The experiments", "also show how the learned object representations can be useful to demonstrate tasks such as writing or pick-and-place tasks, e.g. to make robot learning more data-efficient.", "This paper presented NEMO, a novel approach to unsupervised object detection from short videos of moving objects and negative videos of scenes without those objects.", "By demonstrating data-efficient and robust object detection without the use of image labels, this paper opens up new research directions.", "There are a number of extensions that would improve the presented approach.", "Combining it with ensemble methods, for example, could provide an uncertainty estimate required to infer whether the object is visible in the current frame.", "Integrating the approach with tracking or filtering could exploit temporal consistency not only during training but also during inference.", "For learning multiple objects in the same scene, merging the different object detectors into a single network could improve performance by sharing intermediate features.", "And creating a large-scale data-set for this approach would be very valuable to develop it further.Taking a broader view, the presented approach takes a step towards unsupervised learning of object-based representations.", "While this paper used manually recorded videos, the method can also be applied to data collected by a robot similar to BID13 and BID28 to learn objects autonomously.", "Acquiring such object-based representations could build a bridge to geometric and symbolic reasoning and enable efficient learning, communication, prediction, and planning in object-based representations." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.19354838132858276, 0.11764705181121826, 0.05405404791235924, 0.08695651590824127, 0.0555555522441864, 0.10256409645080566, 0.3333333432674408, 0.0952380895614624, 0.11764705181121826, 0.1428571343421936, 0.20000000298023224, 0.11764705181121826, 0.1875, 0.08695651590824127, 0.09756097197532654, 0.1538461446762085, 0.09302325546741486, 0.4516128897666931, 0.2222222238779068, 0.12903225421905518, 0.0476190447807312, 0.072727270424366, 0.08163265138864517, 0.0833333283662796, 0.08695651590824127, 0.17777776718139648, 0.05714285373687744, 0.04999999701976776, 0.052631575614213943, 0.29411762952804565, 0.1875, 0.0833333283662796, 0.05714285373687744, 0, 0.05714285373687744, 0.09756097197532654, 0.10526315122842789, 0.0624999962747097 ]

[ "Learning to detect objects without image labels from 3 minutes of video" ]

[ "Recent powerful pre-trained language models have achieved remarkable performance on most of the popular datasets for reading comprehension.", "It is time to introduce more challenging datasets to push the development of this field towards more comprehensive reasoning of text.", "In this paper, we introduce a new Reading Comprehension dataset requiring logical reasoning (ReClor) extracted from standardized graduate admission examinations.", "As earlier studies suggest, human-annotated datasets usually contain biases, which are often exploited by models to achieve high accuracy without truly understanding the text.", "In order to comprehensively evaluate the logical reasoning ability of models on ReClor, we propose to identify biased data points and separate them into EASY set while the rest as HARD set.", "Empirical results show that the state-of-the-art models have an outstanding ability to capture biases contained in the dataset with high accuracy on EASY set.", "However, they struggle on HARD set with poor performance near that of random guess, indicating more research is needed to essentially enhance the logical reasoning ability of current models.", "Machine reading comprehension (MRC) is a fundamental task in Natural Language Processing, which requires models to understand a body of text and answer a particular question related to the context.", "With success of unsupervised representation learning in NLP, language pre-training based models such as GPT-2 (Radford et al., 2019) , BERT (Devlin et al., 2019) , XLNet (Yang et al., 2019) and RoBERTa (Liu et al., 2019) have achieved nearly saturated performance on most of the popular MRC datasets (Rajpurkar et al., 2016; Lai et al., 2017; Rajpurkar et al., 2018; Wang et al., 2018) .", "It is time to challenge state-of-the-art models with more difficult reading comprehension tasks and move a step forward to more comprehensive analysis and reasoning over text (Dua et al., 2019) .", "In natural language understanding, logical reasoning is an important ability to examine, analyze and critically evaluate arguments as they occur in ordinary language according to the definition from Law School Admission Council (2019a).", "It is a significant component of human intelligence and is essential in negotiation, debate and writing etc.", "However, existing reading comprehension datasets have none or merely a small amount of data requiring logical reasoning, e.g., 0% in MCTest dataset (Richardson et al., 2013 ) and 1.2% in SQuAD (Rajpurkar et al., 2016) according to Sugawara & Aizawa (2016) .", "One related task is natural language inference, which requires models to label the logical relationships of sentence pairs.", "However, this task only considers three types of simple logical relationships and only needs reasoning at sentence-level.", "To push the development of models in logical reasoning from simple logical relationship classification to multiple complicated logical reasoning and from sentence-level to passage-level, it is necessary to introduce a reading comprehension dataset targeting logical reasoning.", "A typical example of logical reasoning questions is shown in Table 1 .", "Similar to the format of multiple-choice reading comprehension datasets (Richardson et al., 2013; Lai et al., 2017) , it contains a context, a question and four options with only one right answer.", "To answer the question in this example, readers need to identify the logical connections between the lines to pinpoint the conflict, then understand each of the options and select an option that solves the conflict.", "Human minds need extensive training and practice to get used to complex reasoning, and it will take immense efforts for crowdsourcing workers to design such logical reasoning questions.", "Inspired by the datasets extracted from standardized examinations (Lai et al., 2017; Clark et al., 2018) , we build a dataset by selecting such logical reasoning questions from standardized exams such as GMAT 1 and LSAT 2 .", "We finally collect 6,139 pieces of logical reasoning questions, which constitute a Reading Comprehension dataset requiring logical reasoning (ReClor).", "Human-annotated datasets usually contain biases (Schwartz et al., 2017; Cai et al., 2017; Bugert et al., 2017; Poliak et al., 2018; Gururangan et al., 2018; Zellers et al., 2019) , which are often exploited by neural network models as shortcut solutions to achieve high testing accuracy.", "For data points whose options can be selected correctly without knowing the contexts and questions, we classify them as biased ones.", "In order to fully assess the logical reasoning ability of the models, we propose to identify the biased data points and group them as EASY set, and put the rest into HARD set.", "Based on our experiments on these separate sets, we find that even the state-of-the-art models can only perform well on EASY set and struggle on HARD set as shown in Figure 1 .", "This phenomenon shows that current models can well capture the biases in the dataset but lack the ability to understand the text and reason based on connections between the lines.", "On the other hand, human beings perform similarly on both the EASY and HARD set.", "It is thus observed that there is still a long way to go to equip models with true logical reasoning ability.", "The contributions of our paper are two-fold.", "First, we introduce ReClor, a new reading comprehension dataset requiring logical reasoning.", "We use option-only-input baselines trained with different random seeds to identify the data points with biases in the testing set, and group them as EASY set, with the rest as HARD set to facilitate comprehensive evaluation.", "Second, we evaluate several stateof-the-art models on ReClor and find these pre-trained language models can perform well on EASY set but struggle on the HARD set.", "This indicates although current models are good at exploiting biases in the dataset, they are far from capable of performing real logical reasoning yet.", "In this paper, we introduce ReClor, a reading comprehension dataset requiring logical reasoning, with the aim to push research progress on logical reasoning in NLP forward from sentence-level to passage-level and from simple logical reasoning to multiple complicated one.", "We propose to identify biased data points and split the testing set into EASY and HARD group for biased and non-biased data separately.", "We further empirically study the different behaviors of state-of-the-art models on these two testing sets, and find recent powerful transformer-based pre-trained language models have an excellent ability to exploit the biases in the dataset but have difficulty in understanding and reasoning given the non-biased data with low performance close to or slightly better than random guess.", "These results show there is a long way to equip deep learning models with real logical reasoning abilities.", "We hope this work would inspire more research in future to adopt similar split technique and evaluation scheme when reporting their model performance.", "We also show by first fine-tuning on a large-scale dataset RACE then fine-tuning on ReClor, the models could obtain significant improvement, showing the potential of transfer learning to solve reasoning tasks." ]

[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]

[ 0.22727271914482117, 0.13636362552642822, 0.260869562625885, 0.03999999538064003, 0.2181818187236786, 0.20408162474632263, 0.4444444477558136, 0.22641508281230927, 0.1111111044883728, 0.29629629850387573, 0.10526315122842789, 0.1463414579629898, 0.2686567008495331, 0.13636362552642822, 0.1904761791229248, 0.37037035822868347, 0.15789473056793213, 0.2181818187236786, 0.145454540848732, 0.15686273574829102, 0.17543859779834747, 0.3255814015865326, 0.03389830142259598, 0.04255318641662598, 0.14814814925193787, 0.2222222238779068, 0.19230768084526062, 0.04999999701976776, 0.2666666507720947, 0.060606058686971664, 0.4736841917037964, 0.145454540848732, 0.1666666567325592, 0.2448979616165161, 0.4067796468734741, 0.08888888359069824, 0.3243243098258972, 0.27272728085517883, 0.12244897335767746, 0.25925925374031067 ]

[ "We introduce ReClor, a reading comprehension dataset requiring logical reasoning, and find that current state-of-the-art models struggle with real logical reasoning with poor performance near that of random guess." ]