Search ICLR 2019

Searching papers submitted to ICLR 2019 can be painful. You might want to know which paper uses technique X, dataset D, or cites author ME. Unfortunately, search is limited to titles, abstracts, and keywords, missing the actual contents of the paper. This Frankensteinian search has returned from 2018 to help scour the papers of ICLR by ripping out their souls using pdftotext.

Good luck! Warranty's not included :)

Need random search inspiration..? Grab something from the list of all tags! ^_^
How about: maximum entropy rl, stochastic optimization, gcnns, similarity search, dynamic processes ..?

Sanity Disclaimer: As you stare at the continuous stream of ICLR and arXiv papers, don't lose confidence or feel overwhelmed. This isn't a competition, it's a search for knowledge. You and your work are valuable and help carve out the path for progress in our field :)

"Random selection" has 100 results

Overcoming the Disentanglement vs Reconstruction Trade-off via Jacobian Supervision

tl;dr A method to learn image representations that are good at both disentangling factors of variation and obtaining faithful reconstructions.

Learning image representations where the factors of variation are disentangled is typically achieved with an encoder-decoder architecture where a subset of the latent variables is constrained to correspond to specific factors, and the rest of them are considered nuisance variables. This widely used approach has an important drawback: as the dimension of the nuisance variables is increased, better image reconstruction is achieved, but the decoder has the flexibility to ignore the specified factors, thus losing the ability to condition the output on those factors. In this work, we propose to overcome this trade-off by progressively growing the dimension of the latent code, while constraining the Jacobian of the output image with respect to the disentangled variables to remain the same. As a result, the obtained models are effective at both disentangling and reconstruction. We demonstrate the aplicability of this method in both unsupervised and supervised scenarios for learning disentangled representations. In a facial attribute manipulation task, we obtain high quality image generation while smoothly controlling dozens of attributes with a single model. This is an order of magnitude more disentangled factors than state-of-the-art methods, while obtaining visually similar or superior results, and avoiding adversarial training

DeepOBS: A Deep Learning Optimizer Benchmark Suite

tl;dr We provide a software package that drastically simplifies, automates, and improves the evaluation of deep learning optimizers.

Because the choice and tuning of the optimizer affects the speed, and ultimately the performance of deep learning, there is significant past and recent research in this area. Yet, perhaps surprisingly, there is no generally agreed-upon protocol for the quantitative and reproducible evaluation of optimization strategies for deep learning. We suggest routines and benchmarks for stochastic optimization, with special focus on the unique aspects of deep learning, such as stochasticity, tunability and generalization. As the primary contribution, we present DeepOBS, a Python package of deep learning optimization benchmarks. The package addresses key challenges in the quantitative assessment of stochastic optimizers, and automates most steps of benchmarking. The library includes a wide and extensible set of ready-to-use realistic optimization problems, such as training Residual Networks for image classification on ImageNet or character-level language prediction models, as well as popular classics like MNIST and CIFAR-10. The package also provides realistic baseline results for the most popular optimizers on these test problems, ensuring a fair comparison to the competition when benchmarking new optimizers, and without having to run costly experiments. It comes with output back-ends that directly produce LaTeX code for inclusion in academic publications. It is written in TensorFlow and available open source.

Universal Successor Features for Transfer Reinforcement Learning

No tl;dr =[

Transfer in Reinforcement Learning (RL) refers to the idea of applying knowledge gained from previous tasks to solving related tasks. Learning a universal value function (Schaul et al., 2015), which generalizes over goals and states, has previously been shown to be useful for transfer. However, successor features are believed to be more suitable than values for transfer (Dayan, 1993; Barreto et al.,2017), even though they cannot directly generalize to new goals. In this paper, we propose (1) Universal Successor Features (USFs) to capture the underlying dynamics of the environment while allowing generalization to unseen goals and (2) a flexible end-to-end model of USFs that can be trained by interacting with the environment. We show that learning USFs is compatible with any RL algorithm that learns state values using a temporal difference method. Our experiments in a simple gridworld and with two MuJoCo environments show that USFs can greatly accelerate training when learning multiple tasks and can effectively transfer knowledge to new tasks.

Meta Learning with Fast/Slow Learners

tl;dr We applied multiple meta-strategy to improve meta-learning performance on base CNNs.

Meta-learning has recently achieved success in many optimization problems. In general, a meta learner g(.) could be learned for a base model f(.) on a variety of tasks, such that it can be more efficient on a new task. In this paper, we make some key modifications to enhance the performance of meta-learning models. (1) we leverage different meta-strategies for different modules to optimize them separately: we use conservative “slow learners” on low-level basic feature representation layers and “fast learners” on high-level task-specific layers; (2) Furthermore, we provide theoretical analysis on why the proposed approach works, based on a case study on a two-layer MLP. We evaluate our model on synthetic MLP regression, as well as low-shot learning tasks on Omniglot and ImageNet benchmarks. We demonstrate that our approach is able to achieve state-of-the-art performance.

Unsupervised classification into unknown k classes

No tl;dr =[

We propose a novel spectral decomposition framework for the unsupervised classification task. Unlike the widely used classification method, this architecture does not require the labels of data and the number of classes. Our key idea is to introduce a piecewise linear map and a spectral decomposition method on the dimension reduced space into generative adversarial networks. Inspired by the human visual recognition system, the proposed framework can classify and also generate images as the human brains do. We build a piecewise linear connection analogous to the cerebral cortex, between the discriminator D and the generator G. This connection allows us to estimate the number of classes k and extract the vectors that represent each class. We show that our framework has the reasonable performance in the experiment.

Accumulation Bit-Width Scaling For Ultra-Low Precision Training Of Deep Networks

tl;dr We present an analytical framework to determine accumulation bit-width requirements in all three deep learning training GEMMs and verify the validity and tightness of our method via benchmarking experiments.

Efforts to reduce the numerical precision of computations in deep learning training have yielded systems that aggressively quantize weights and activations, yet employ wide high-precision accumulators for partial sums in inner-product operations to preserve the quality of convergence. The absence of any framework to analyze the precision requirements of partial sum accumulations results in conservative design choices. This imposes an upper-bound on the reduction of complexity of multiply-accumulate units. We present a statistical approach to analyze the impact of reduced accumulation precision on deep learning training. Observing that a bad choice for accumulation precision results in loss of information that manifests itself as a reduction in variance in an ensemble of partial sums, we derive a set of equations that relate this variance to the length of accumulation and the minimum number of bits needed for accumulation. We apply our analysis to three benchmark networks: CIFAR-10 ResNet 32, ImageNet ResNet 18 and ImageNet AlexNet. In each case, with accumulation precision set in accordance with our proposed equations, the networks successfully converge to the single precision floating-point baseline. We also show that reducing accumulation precision further degrades the quality of the trained network, proving that our equations produce tight bounds. Overall this analysis enables precise tailoring of computation hardware to the application, yielding area- and power-optimal systems.

No tl;dr =[

Being able to predict what may happen in the future requires an in-depth understanding of the physical and causal rules that govern the world. A model that is able to do so has a number of appealing applications, from robotic planning to representation learning. However, learning to predict raw future observations, such as frames in a video, is exceedingly challenging—the ambiguous nature of the problem can cause a naively designed model to average together possible futures into a single, blurry prediction. Recently, this has been addressed by two distinct approaches: (a) latent variational variable models that explicitly model underlying stochasticity and (b) adversarially-trained models that aim to produce naturalistic images. However, a standard latent variable model can struggle to produce realistic results, and a standard adversarially-trained model underutilizes latent variables and fails to produce diverse predictions. We show that these distinct methods are in fact complementary. Combining the two produces predictions that look more realistic to human raters and better cover the range of possible futures. Our method outperforms prior works in these aspects.

HAPPIER: Hierarchical Polyphonic Music Generative RNN

No tl;dr =[

Generating polyphonic music with coherent global structure is a major challenge for automatic composition algorithms. The primary difficulty arises due to the inefficiency of models to recognize underlying patterns beneath music notes across different levels of time scales and remain long-term consistency while composing. Hierarchical architectures can capture and represent learned patterns in different temporal scales and maintain consistency over long time spans, and this corresponds to the hierarchical structure in music. Motivated by this, focusing on leveraging the idea of hierarchical models and improve them to fit the sequence modeling problem, our paper proposes HAPPIER: a novel HierArchical PolyPhonic musIc gEnerative RNN. In HAPPIER, A higher measure level' learns correlations across measures and patterns for chord progressions, and a lower note level' learns a conditional distribution over the notes to generate within a measure. The two hierarchies operate at different clock rates: the higher one operates on a longer timescale and updates every measure, while the lower one operates on a shorter timescale and updates every unit duration. The two levels communicate with each other, and thus the entire architecture is trained jointly end-to-end by back-propagation. HAPPIER, profited from the strength of the hierarchical structure, generates polyphonic music with long-term dependencies compared to the state-of-the-art methods.

Graph U-Net

tl;dr We propose the graph U-Net based on our novel graph pooling and unpooling layer for network embedding.

We consider the problem of representation learning for graph data. Convolutional neural networks can naturally operate on images, but have significant challenges in dealing with graph data. Given images are special cases of graphs with nodes lie on 2D lattices, graph embedding tasks have a natural correspondence with image pixel-wise prediction tasks such as segmentation. While encoder-decoder architectures like U-Net have been successfully applied on many image pixel-wise prediction tasks, similar methods are lacking for graph data. This is due to the fact that pooling and up-sampling operations are not natural on graph data. To address these challenges, we propose novel graph pooling (gPool) and unpooling (gUnpool) operations in this work. The gPool layer adaptively selects some nodes to form a smaller graph based on their scalar projection values on a trainable projection vector. We further propose the gUnpool layer as the inverse operation of the gPool layer. The gUnpool layer restores the graph into its original structure using the position information of nodes selected in the corresponding gPool layer. Based on our proposed gPool and gUnpool layers, we develop an encoder-decoder model on graph, known as the graph U-Net. Our experimental results on node classification tasks demonstrate that our methods achieve consistently better performance than previous models.

Global-to-local Memory Pointer Networks for Task-Oriented Dialogue

tl;dr We propose a global memory encoder and a global memory decoder that share an external knowledge to strengthen task-oriented dialogue generation via sketch responses and pointer networks.

End-to-end task-oriented dialogue is challenging since knowledge bases are usually large, dynamic and hard to incorporate into a learning framework. We propose the global-to-local memory pointer (GLMP) networks to address this issue. In our model, a global memory encoder and a local memory decoder are proposed to share an external knowledge. The encoder encodes dialogue history, modifies global contextual representation that is shared with the decoder, and generates a global memory pointer. The decoder first generates a sketch response with unfilled slots. Next, it passes the global memory pointer to filter the external knowledge for relevant information, then instantiates the slots via the local memory pointers which points to specific entries in the external knowledge. We empirically show that our model can improve copy accuracy and mitigate the common out-of-vocabulary problem. As a result, GLMP is able to improve over the previous state-of-the-art models in both simulated bAbI Dialogue and human-human Stanford Multi-domain Dialogue datasets on automatic and human evaluation.

A Data-Driven and Distributed Approach to Sparse Signal Representation and Recovery

tl;dr We use deep learning techniques to solve the sparse signal representation and recovery problem.

In this paper, we focus on two challenges which offset the promise of sparse signal representation, sensing, and recovery. First, real-world signals can seldom be described as perfectly sparse vectors in a known basis, and traditionally used random measurement schemes are seldom optimal for sensing them. Second, existing signal recovery algorithms are usually not fast enough to make them applicable to real-time problems. In this paper, we address these two challenges by presenting a novel framework based on deep learning. For the first challenge, we cast the problem of finding informative measurements by using a maximum likelihood (ML) formulation and show how we can build a data-driven dimensionality reduction protocol for sensing signals using convolutional architectures. For the second challenge, we discuss and analyze a novel parallelization scheme and show it significantly speeds-up the signal recovery process. We demonstrate the significant improvement our method obtains over competing methods through a series of experiments.

COMPOSITION AND DECOMPOSITION OF GANS

tl;dr GANs can be composed to build more complex models and decomposed to obtain building blocks

In this work, we propose a composition/decomposition framework for adversarially training generative models on composed data - data where each sample can be thought of as being constructed from a fixed number of components. In our framework, samples are generated by sampling components from component generators and feeding these components to a composition function which combines them into a “composed sample”. This compositional training approach improves the modularity, extensibility and interpretability of Generative Adversarial Networks (GANs) - providing a principled way to incrementally construct complex models out of simpler component models, and allowing for explicit “division of responsibility” between these components. Using this framework, we define a family of learning tasks and evaluate their feasibility on two datasets in two different data modalities (image and text). Lastly, we derive sufficient conditions such that these compositional generative models are identifiable. Our work provides a principled approach to building on pretrained generative models or for exploiting the compositional nature of data distributions to train extensible and interpretable models.

Learning Localized Generative Models for 3D Point Clouds via Graph Convolution

tl;dr A GAN using graph convolution operations with dynamically computed graphs from hidden features

Point clouds are an important type of geometric data and have widespread use in computer graphics and vision. However, learning representations for point clouds is particularly challenging due to their nature as being an unordered collection of points irregularly distributed in 3D space. Graph convolution, a generalization of the convolution operation for data defined over graphs, has been recently shown to be very successful at extracting localized features from point clouds in supervised or semi-supervised tasks such as classification or segmentation. This paper studies the unsupervised problem of a generative model exploiting graph convolution. We focus on the generator of a GAN and define methods for graph convolution when the graph is not known in advance as it is the very output of the generator. The proposed architecture learns to generate localized features that approximate graph embeddings of the output geometry. We also study the problem of defining an upsampling layer in the graph-convolutional generator, whereby it learns to exploit a self-similarity prior to sample the data distribution.

RelGAN: Relational Generative Adversarial Networks for Text Generation

No tl;dr =[

Generative adversarial networks (GANs) have achieved great success at generating realistic images. However, the text generation still remains a challenging task for modern GAN architectures. In this work, we propose RelGAN, a new GAN architecture for text generation, consisting of three main components: a relational memory based generator for the long-distance dependency modeling, the Gumbel-Softmax relaxation for training GANs on discrete data, and multiple embedded representations in the discriminator to provide a more informative signal for the generator updates. Our experiments show that RelGAN outperforms current state-of-the-art models in terms of sample quality and diversity, and we also reveal via ablation studies that each component of RelGAN contributes critically to its performance improvements. Moreover, a key advantage of our method, that distinguishes it from other GANs, is the ability to control the trade-off between sample quality and diversity via the use of a single adjustable parameter. Finally, RelGAN is the first architecture that makes GANs with Gumbel-Softmax relaxation succeed in generating realistic text.

Discrete Structural Planning for Generating Diverse Translations

tl;dr Learning discrete structural representation to control sentence generation and obtain diverse outputs

Planning is important for humans when producing complex languages, which is a missing part in current language generation models. In this work, we add a planning phase in neural machine translation to control the global sentence structure ahead of translation. Our approach learns discrete structural representations to encode syntactic information of target sentences. During translation, we can either let beam search to choose the structural codes automatically or specify the codes manually. The word generation is then conditioned on the selected discrete codes. Experiments show that the translation performance remains intact by learning the codes to capture pure structural variations. Through structural planning, we are able to control the global sentence structure by manipulating the codes. By evaluating with a proposed structural diversity metric, we found that the sentences sampled using different codes have much higher diversity scores. In qualitative analysis, we demonstrate that the sampled paraphrase translations have drastically different structures.

Invariant-covariant representation learning

tl;dr This paper presents a novel latent-variable generative modelling technique that enables the representation of global information into one latent variable and local information into another latent variable.

Representations learnt through deep neural networks tend to be highly informative, but opaque in terms of what information they learn to encode. We introduce an approach to probabilistic modelling that learns to represent data with two separate deep representations: an invariant representation that encodes the information of the class from which the data belongs, and a covariant representation that encodes the symmetry transformation defining the particular data point within the class manifold (covariant in the sense that the representation varies naturally with symmetry transformations). This approach to representation learning is conceptually transparent, easy to implement, and in-principle generally applicable to any data comprised of discrete classes of continuous distributions (e.g. objects in images, topics in language, individuals in behavioural data). We demonstrate qualitatively compelling representation learning and competitive quantitative performance, in both supervised and semi-supervised settings, versus comparable modelling approaches in the literature with little fine tuning.

ALISTA: Analytic Weights Are As Good As Learned Weights in LISTA

No tl;dr =[

Deep neural networks based on unfolding an iterative algorithm, for example, LISTA (learned iterative shrinkage thresholding algorithm), have been an empirical success for sparse signal recovery. The weights of these neural networks are currently determined by data-driven “black-box” training. In this work, we propose Analytic LISTA (ALISTA), where the weight matrix in LISTA is computed as the solution to a data-free optimization problem, leaving only the stepsize and threshold parameters to data-driven learning. This signiﬁcantly simpliﬁes the training. Speciﬁcally, the data-free optimization problem is based on coherence minimization. We show our ALISTA retains the optimal linear convergence proved in (Chen et al., 2018) and has a performance comparable to LISTA. Furthermore, we extend ALISTA to convolutional linear operators, again determined in a data-free manner. We also propose a feed-forward framework that combines the data-free optimization and ALISTA networks from end to end, one that can be jointly trained to gain robustness to small perturbations in the encoding model.

Discriminative out-of-distribution detection for semantic segmentation

tl;dr We present a novel approach for detecting out-of-distribution pixels in semantic segmentation.

Most classification and segmentation datasets assume a closed-world scenario in which predictions are expressed as distribution over a predetermined set of visual classes. However, such assumption implies unavoidable and often unnoticeable failures in presence of out-of-distribution (OOD) input. These failures are bound to happen in most real-life applications since current visual ontologies are far from being comprehensive. We propose to address this issue by discriminative detection of OOD pixels in input data. Different from recent approaches, we avoid to bring any decisions by only observing the training dataset of the primary model trained to solve the desired computer vision task. Instead, we train a dedicated OOD model which discriminates the primary training set from a much larger "background" dataset which approximates the variety of the visual world. We perform our experiments on high resolution natural images in a dense prediction setup. We use several road driving datasets as our training distribution, while we approximate the background distribution with the ILSVRC dataset. We evaluate our approach on WildDash test, which is currently the only public test dataset with out-of-distribution images. The obtained results show that the proposed approach succeeds to identify out-of-distribution pixels while outperforming previous work by a wide margin.

Analysis of Memory Organization for Dynamic Neural Networks

No tl;dr =[

An increasing number of neural memory networks have been developed, leading to the need for a systematic approach to analyze and compare their underlying memory structures. Thus, in this paper, we first create a framework for memory organization and then compare four popular dynamic models: vanilla recurrent neural network, long short term memory, neural stack and neural RAM. This analysis helps to open the dynamic neural network' black box from the memory usage prospective. Accordingly, a taxonomy for these networks and their variants is proposed and proved using a unifying architecture. With the taxonomy, both network architectures and learning tasks are classified into four classes. And a one-to-one mapping is built between them to help practitioners select the appropriate architecture. To exemplify each task type, four synthetic tasks with different memory requirements are developed. Moreover, we use two natural language processing applications to apply the methodology in a realistic setting.

Marginal Policy Gradients: A Unified Family of Estimators for Bounded Action Spaces with Applications

No tl;dr =[

Many complex domains, such as robotics control and real-time strategy (RTS) games, require an agent to learn a continuous control. In the former, an agent learns a policy over R^d and in the latter, over a discrete set of actions each of which is parametrized by a continuous parameter. Such problems are naturally solved using policy based reinforcement learning (RL) methods, but unfortunately these often suffer from high variance leading to instability and slow convergence. Unnecessary variance is introduced whenever policies over bounded action spaces are modeled using distributions with unbounded support by applying a transformation T to the sampled action before execution in the environment. Recently, the variance reduced clipped action policy gradient (CAPG) was introduced for actions in bounded intervals, but to date no variance reduced methods exist when the action is a direction, something often seen in RTS games. To this end we introduce the angular policy gradient (APG), a stochastic policy gradient method for directional control. With the marginal policy gradients family of estimators we present a unified analysis of the variance reduction properties of APG and CAPG; our results provide a stronger guarantee than existing analyses for CAPG. Experimental results on a popular RTS game and a navigation task show that the APG estimator offers a substantial improvement over the standard policy gradient.

An investigation of model-free planning

No tl;dr =[

The field of reinforcement learning (RL) is facing increasingly challenging domains with combinatorial complexity. For an RL agent to address these challenges, it is essential that it can plan effectively. Prior work has typically utilized an explicit model of the environment, combined with a specific planning algorithm (such as tree search). More recently, a new family of methods have been proposed that learn how to plan, by providing the structure for planning via an inductive bias in the function approximator (such as a tree structured neural network), trained end-to-end by a model-free RL algorithm. In this paper, we go even further, and suggest that an entirely model-free approach, without any special structure beyond standard neural network components such as convolutional networks and LSTMs, can learn to plan effectively. We measure our agent's effectiveness at planning in terms of its ability to generalize across a combinatorial and irreversible state space, its data efficiency, and its ability to utilize additional thinking time. We find that our agent has the characteristics that one might expect to find in a true planning algorithm. Furthermore, it exceeds the state-of-the-art in challenging combinatorial domains such as Sokoban and outperforms other model-free approaches that utilize strong inductive biases towards planning.

Transferring Knowledge across Learning Processes

tl;dr We propose Leap, a framework that transfers knowledge across learning processes by minimizing the expected distance the training process travels on a task's loss surface.

In complex transfer learning scenarios new tasks might not be tightly linked to previous tasks. Approaches that transfer information contained only in the final parameters of a source model will therefore struggle. Instead, transfer learning at a higher level of abstraction is needed. We propose Leap, a framework that achieves this by transferring knowledge across learning processes. We associate each task with a manifold on which the training process travels from initialization to final parameters and construct a meta learning objective that minimizes the expected length of this path. Our framework leverages only information obtained during training and can be computed on the fly at negligible cost. We demonstrate that our framework outperforms competing methods, both in meta learning and transfer learning, on a set of computer vision tasks. Finally, we demonstrate that Leap can transfer knowledge across learning processes in demanding Reinforcement learning environments (Atari) that involve millions of gradient steps.

Interpretable Continual Learning

tl;dr The paper develops an interpretable continual learning framework where explanations of the finished tasks are used to enhance the attention of the learner during the future tasks, and where an explanation metric is proposed too.

We present a framework for interpretable continual learning (ICL). We show that explanations of previously performed tasks can be used to improve performance on future tasks. ICL generates a good explanation of a finished task, then uses this to focus attention on what is important when facing a new task. The ICL idea is general and may be applied to many continual learning approaches. Here we focus on the variational continual learning framework to take advantage of its flexibility and efficacy in overcoming catastrophic forgetting. We use saliency maps to provide explanations of performed tasks and propose a new metric to assess their quality. Experiments show that ICL achieves state-of-the-art results in terms of overall continual learning performance as measured by average classification accuracy, and also in terms of its explanations, which are assessed qualitatively and quantitatively using the proposed metric.

Convolutional CRFs for Semantic Segmentation

tl;dr We propose Convolutional CRFs a fast, powerful and trainable alternative to Fully Connected CRFs.

For the challenging semantic image segmentation task the best performing models have traditionally combined the structured modelling capabilities of Conditional Random Fields (CRFs) with the feature extraction power of CNNs. In more recent works however, CRF post-processing has fallen out of favour. We argue that this is mainly due to the slow training and inference speeds of CRFs, as well as the difficulty of learning the internal CRF parameters. To overcome both issues we propose to add the assumption of conditional independence to the framework of fully-connected CRFs. This allows us to reformulate the inference in terms of convolutions, which can be implemented highly efficiently on GPUs.Doing so speeds up inference and training by two orders of magnitude. All parameters of the convolutional CRFs can easily be optimized using backpropagation. Towards the goal of facilitating further CRF research we have made our implementations publicly available.

Learning Goal-Conditioned Value Functions with one-step Path rewards rather than Goal-Rewards

tl;dr Do Goal-Conditioned Value Functions need Goal-Rewards to Learn?

Multi-goal reinforcement learning (MGRL) addresses tasks where the desired goal state can change for every trial. State-of-the-art algorithms model these problems such that the reward formulation depends on the goals, to associate them with high reward. This dependence introduces additional goal reward resampling steps in algorithms like Hindsight Experience Replay (HER) that reuse trials in which the agent fails to reach the goal by recomputing rewards as if reached states were psuedo-desired goals. We propose a reformulation of goal-conditioned value functions for MGRL that yields a similar algorithm, while removing the dependence of reward functions on the goal. Our formulation thus obviates the requirement of reward-recomputation that is needed by HER and its extensions. We also extend a closely related algorithm, Floyd-Warshall Reinforcement Learning, from tabular domains to deep neural networks for use as a baseline. Our results are competetive with HER while substantially improving sampling efficiency in terms of reward computation.

Multi-way Encoding for Robustness to Adversarial Attacks

tl;dr We demonstrate that by leveraging a multi-way output encoding, rather than the widely used one-hot encoding, we can make deep models more robust to adversarial attacks.

Deep models are state-of-the-art for many computer vision tasks including image classification and object detection. However, it has been shown that deep models are vulnerable to adversarial examples. We highlight how one-hot encoding directly contributes to this vulnerability and propose breaking away from this widely-used, but highly-vulnerable mapping. We demonstrate that by leveraging a different output encoding, multi-way encoding, we can make models more robust. Our approach makes it more difficult for adversaries to find useful gradients for generating adversarial attacks. We present state-of-the-art robustness results for black-box, white-box attacks, and achieve higher clean accuracy on four benchmark datasets: MNIST, CIFAR-10, CIFAR-100, and SVHN when combined with adversarial training. The strength of our approach is also presented in the form of an attack for model watermarking, raising challenges in detecting stolen models.

Context Dependent Modulation of Activation Function

tl;dr We propose a modification to traditional Artificial Neural Networks motivated by the biology of neurons to enable the shape of the activation function to be context dependent.

We propose a modification to traditional Artificial Neural Networks (ANNs), which provides the ANNs with new aptitudes motivated by biological neurons. Biological neurons work far beyond linearly summing up synaptic inputs and then transforming the integrated information. A biological neuron change firing modes accordingly to peripheral factors (e.g., neuromodulators) as well as intrinsic ones. Our modification connects a new type of ANN nodes, which mimic the function of biological neuromodulators and are termed modulators, to enable other traditional ANN nodes to adjust their activation sensitivities in run-time based on their input patterns. In this manner, we enable the slope of the activation function to be context dependent. This modification produces statistically significant improvements in comparison with traditional ANN nodes in the context of Convolutional Neural Networks and Long Short-Term Memory networks.

The role of over-parametrization in generalization of neural networks

tl;dr We suggest a generalization bound that could potentially explain the improvement in generalization with over-parametrization.

Despite existing work on ensuring generalization of neural networks in terms of scale sensitive complexity measures, such as norms, margin and sharpness, these complexity measures do not offer an explanation of why neural networks generalize better with over-parametrization. In this work we suggest a novel complexity measure based on unit-wise capacities resulting in a tighter generalization bound for two layer ReLU networks. Our capacity bound correlates with the behavior of test error with increasing network sizes, and could potentially explain the improvement in generalization with over-parametrization. We further present a matching lower bound for the Rademacher complexity that improves over previous capacity lower bounds for neural networks.

Selective Convolutional Units: Improving CNNs via Channel Selectivity

tl;dr We propose a new module that improves any ResNet-like architectures by enforcing "channel selective" behavior to convolutional layers

Bottleneck structures with identity (e.g., residual) connection are now emerging popular paradigms for designing deep convolutional neural networks (CNN), for processing large-scale features efficiently. In this paper, we focus on the information-preserving nature of the bottleneck structures and utilize this to enable a convolutional layer to have a new functionality of channel-selectivity, i.e., focusing its computations on important channels. In particular, we propose Selective Convolutional Unit (SCU), an easy-to-use architectural unit that improves parameter efficiency of various modern CNNs with bottlenecks. During training, SCU gradually learns the channel-selectivity on-the-fly via the alternative usage of (a) pruning unimportant channels, and (b) rewiring the pruned parameters to important channels. The rewired parameters emphasize the target channel in a way that selectively enlarges the convolutional kernels corresponding to it. Our experimental results demonstrate that the SCU-based models without any post-processing generally achieve both model compression and accuracy improvement compared to the baselines, consistently for all tested architectures.

COLLABORATIVE MULTIAGENT REINFORCEMENT LEARNING IN HOMOGENEOUS SWARMS

tl;dr Novel policy gradient for multiagent systems via distributed learning.

A deep reinforcement learning solution is developed for a collaborative multiagent system. Individual agents choose actions in response to the state of the environment, their own state, and possibly partial information about the state of other agents. Actions are chosen to maximize a collaborative long term discounted reward that encompasses the individual rewards collected by each agent. The paper focuses on developing a scalable approach that applies to large swarms of homogeneous agents. This is accomplished by forcing the policies of all agents to be the same resulting in a constrained formulation in which the experiences of each agent inform the learning process of the whole team, thereby enhancing the sample efficiency of the learning process. A projected coordinate policy gradient descent algorithm is derived to solve the constrained reinforcement learning problem. Experimental evaluations in collaborative navigation, a multi-predator-multi-prey game, and a multiagent survival game show marked improvements relative to methods that do not exploit the policy equivalence that naturally arises in homogeneous swarms.

Teacher Guided Architecture Search

tl;dr Faster architecture search by maximizing representational similarity with a teacher network

Strong improvements in neural network performance in vision tasks have resulted from the search of alternative network architectures, and prior work has shown that this search process can be automated and guided by evaluating candidate network performance following limited training (“Performance Guided Architecture Search” or PGAS). However, because of the large architecture search spaces and the high computational cost associated with evaluating each candidate model, further gains in computational efficiency are needed. Here we present a method termed Teacher Guided Search for Architectures by Generation and Evaluation (TG-SAGE) that produces up to an order of magnitude in search efficiency over PGAS methods. Specifically, TG-SAGE guides each step of the architecture search by evaluating the similarity of internal representations of the candidate networks with those of the (fixed) teacher network. We show that this procedure leads to significant reduction in required per-sample training and that, this advantage holds for two different search spaces of architectures, and two different search algorithms. We further show that in the space of convolutional cells for visual categorization, TG-SAGE finds a cell structure with similar performance as was previously found using other methods but at a total computational cost that is two orders of magnitude lower than Neural Architecture Search (NAS) and more than four times lower than progressive neural architecture search (PNAS). These results suggest that TG-SAGE can be used to accelerate network architecture search in cases where one has access to some or all of the internal representations of a teacher network of interest, such as the brain.

tl;dr We proposal a simple surrogate for gradient descent to improve training of deep neural nets and other optimization problems.

We propose a class of very simple modifications of gradient descent and stochastic gradient descent. We show that when applied to a large variety of machine learning problems, ranging from softmax regression to deep neural nets, the proposed surrogates can dramatically reduce the variance and improve the generalization accuracy. The methods only involve multiplying the usual (stochastic) gradient by the inverse of a positive definitive matrix coming from the discrete Laplacian or high order generalizations. The theory of Hamilton-Jacobi partial differential equations demonstrates that the new algorithm is almost the same as doing gradient descent on a new function which (i) has the same global minima as the original function and (ii) is “more convex”. We show that optimization algorithms with these surrogates converge uniformly in the discrete Sobolev H^p_\sigma sense and reduce the optimality gap for convex optimization problems. We implement our algorithm into both PyTorch and Tensorflow platforms which only involves changing of a few lines of code. The code will be available on Github.

Function Space Particle Optimization for Bayesian Neural Networks

No tl;dr =[

While Bayesian neural networks (BNNs) have drawn increasing attention, their posterior inference remains challenging, due to the high-dimensional and over-parameterized nature. To address this issue, several highly flexible and scalable variational inference procedures based on the idea of particle optimization have been proposed. These methods directly optimize a set of particles to approximate the target posterior. However, their application to BNNs often yields sub-optimal performance, as such methods have a particular failure mode on over-parameterized models. In this paper, we propose to solve this issue by performing particle optimization directly in the space of regression functions. We demonstrate through extensive experiments that our method successfully overcomes this issue, and outperforms strong baselines in a variety of tasks including prediction, defense against adversarial examples, and reinforcement learning.

Harmonic Unpaired Image-to-image Translation

tl;dr Smooth regularization over sample graph for unpaired image-to-image translation results in significantly improved consistency

The recent direction of unpaired image-to-image translation is on one hand very exciting as it alleviates the big burden in obtaining label-intensive pixel-to-pixel supervision, but it is on the other hand not fully satisfactory due to the presence of artifacts and degenerated transformations. In this paper, we take a manifold view of the problem by introducing a smoothness constraint over the sample graph to attain harmonic functions to enforce consistent mappings during the translation. We develop HarmonicGAN to learn bi-directional translations between the source and the target domain. With the help of similarity-consistency, the inherent self-consistency property of samples can be maintained. Distance metrics defined on two types of features including histogram and CNN are exploited. Under an identical problem setting as CycleGAN without additional manual inputs, HarmonicGAN demonstrates a significant qualitative and quantitative improvement over the state of the art, as well as improved interpretability. We show experimental results in a number of applications including medical imaging, object transfiguration, and semantic labeling. We outperform the competing methods in all tasks, and for a medical imaging task in particular our method turns CycleGAN from a failure to a success, halving the mean-squared error, and generating images that radiologists prefer over competing methods in 95% of cases.

Generalized Capsule Networks with Trainable Routing Procedure

tl;dr A scalable capsule network

CapsNet (Capsule Network) was first proposed by Sabour et al. (2017) and lateranother version of CapsNet was proposed by Hinton et al. (2018). CapsNet hasbeen proved effective in modeling spatial features with much fewer parameters.However, the routing procedures (dynamic routing and EM routing) in both pa-pers are not well incorporated into the whole training process, and the optimalnumber for the routing procedure has to be found manually. We propose Gen-eralized GapsNet (G-CapsNet) to overcome this disadvantages by incorporatingthe routing procedure into the optimization. We implement two versions of G-CapsNet (fully-connected and convolutional) on CAFFE (Jia et al. (2014)) andevaluate them by testing the accuracy on MNIST & CIFAR10, the robustness towhite-box & black-box attack, and the generalization ability on GAN-generatedsynthetic images. We also explore the scalability of G-CapsNet by constructinga relatively deep G-CapsNet. The experiment shows that G-CapsNet has goodgeneralization ability and scalability.

Reconciling Feature-Reuse and Overfitting in DenseNet with Specialized Dropout

tl;dr Realizing the drawbacks when applying original dropout on DenseNet, we craft the design of dropout method from three aspects, the idea of which could also be applied on other CNN models.

Recently convolutional neural networks (CNNs) achieve great accuracy in visual recognition tasks. DenseNet becomes one of the most popular CNN models due to its effectiveness in feature-reuse. However, like other CNN models, DenseNets also face overfitting problem if not severer. Existing dropout method can be applied but not as effective due to the introduced nonlinear connections. In particular, the property of feature-reuse in DenseNet will be impeded, and the dropout effect will be weakened by the spatial correlation inside feature maps. To address these problems, we craft the design of a specialized dropout method from three aspects, dropout location, dropout granularity, and dropout probability. The insights attained here could potentially be applied as a general approach for boosting the accuracy of other CNN models with similar nonlinear connections. Experimental results show that DenseNets with our specialized dropout method yield better accuracy compared to vanilla DenseNet and state-of-the-art CNN models, and such accuracy boost increases with the model depth.

ON RANDOM DEEP AUTOENCODERS: EXACT ASYMPTOTIC ANALYSIS, PHASE TRANSITIONS, AND IMPLICATIONS TO TRAINING

tl;dr We study the behavior of weight-tied multilayer vanilla autoencoders under the assumption of random weights. Via an exact characterization in the limit of large dimensions, our analysis reveals interesting phase transition phenomena.

We study the behavior of weight-tied multilayer vanilla autoencoders under the assumption of random weights. Via an exact characterization in the limit of large dimensions, our analysis reveals interesting phase transition phenomena when the depth becomes large. This, in particular, provides quantitative answers and insights to three questions that were yet fully understood in the literature. Firstly, we provide a precise answer on how the random deep weight-tied autoencoder model performs “approximate inference” as posed by Scellier et al. (2018), and its connection to reversibility considered by several theoretical studies. Secondly, we show that deep autoencoders display a higher degree of sensitivity to perturbations in the parameters, distinct from the shallow counterparts. Thirdly, we obtain insights on pitfalls in training initialization practice, and demonstrate experimentally that it is possible to train a deep autoencoder, even with the tanh activation and a depth as large as 200 layers, without resorting to techniques such as layer-wise pre-training or batch normalization. Our analysis is not specific to any depths or any Lipschitz activations, and our analytical techniques may have broader applicability.

Trellis Networks for Sequence Modeling

tl;dr Trellis networks are a new sequence modeling architecture that bridges recurrent and convolutional models and sets a new state of the art on word- and character-level language modeling.

We present trellis networks, a new architecture for sequence modeling. On the one hand, a trellis network is a temporal convolutional network with special structure, characterized by weight tying across depth and direct injection of the input into deep layers. On the other hand, we show that truncated recurrent networks are equivalent to trellis networks with special sparsity structure in their weight matrices. Thus trellis networks with general weight matrices generalize truncated recurrent networks. We leverage these connections to design high-performing trellis networks that absorb structural and algorithmic elements from both recurrent and convolutional models. Experiments demonstrate that trellis networks outperform the current state of the art on a variety of challenging benchmarks, including word-level language modeling on Penn Treebank and WikiText-103, character-level language modeling on Penn Treebank, and stress tests designed to evaluate long-term memory retention.

Learning powerful policies and better generative models by interaction

tl;dr In this paper, we formulate a way to ensure consistency between the predictions of dynamics model and the real observations from the environment. Thus allowing us to learn powerful policies, as well as better dynamics models.

Model-based reinforcement learning approaches have the promise of being sample efficient. Much of the progress in learning dynamics models in RL has been made by learning models via supervised learning. There is enough evidence that humans build a model of the environment, not only by observing the environment but also by interacting with the environment. Interaction with the environment allows humans to carry out \textit{experiments}: taking actions that help uncover true casual relationships which in turn can be used for building better dynamics models. Analogously, we would expect such interaction to be helpful for a learning agent while it learns to model the dynamics of the environment. In this paper, we build upon this intuition, by using an auxiliary cost function to ensure consistency between what the agent observes (by actually performing actions in the real world) and what it hallucinates (by imagining to have taken actions in the environment). Our empirical analysis shows that the proposed approach helps to train powerful policies as well as better dynamics models.

Gaussian-gated LSTM: Improved convergence by reducing state updates

tl;dr Gaussian-gated LSTM is a novel time-gated LSTM RNN network that enables faster and better training on long sequence data.

Recurrent neural networks can be difficult to train on long sequence data due to the well-known vanishing gradient problem. Some architectures incorporate methods to reduce RNN state updates, therefore allowing the network to preserve memory over long temporal intervals. To address these problems of convergence, this paper proposes a timing-gated LSTM RNN model, called the Gaussian-gated LSTM (g-LSTM). The time gate controls when a neuron can be updated during training, enabling longer memory persistence and better error-gradient flow. This model captures long-temporal dependencies better than an LSTM and the time gate parameters can be learned even from non-optimal initialization values. Because the time gate limits the updates of the neuron state, the number of computes needed for the network update is also reduced. By adding a computational budget term to the training loss, we can obtain a network which further reduces the number of computes by at least 10x. Finally, by employing a temporal curriculum learning schedule for the g-LSTM, we can reduce the convergence time of the equivalent LSTM network on long sequences.

Manifold Alignment via Feature Correspondence

tl;dr We propose a method for aligning the latent features learned from different datasets using harmonic correlations.

We propose a novel framework for combining datasets via alignment of their associated intrinsic dimensions. Our approach assumes that the two datasets are sampled from a common latent space, i.e., they measure equivalent systems. Thus, we expect there to exist a natural (albeit unknown) alignment of the data manifolds associated with the intrinsic geometry of these datasets, which are perturbed by measurement artifacts in the sampling process. Importantly, we do not assume any individual correspondence (partial or complete) between data points. Instead, we rely on our assumption that a subset of data features have correspondence across datasets. We leverage this assumption to estimate relations between intrinsic manifold dimensions, which are given by diffusion map coordinates over each of the datasets. We compute a correlation matrix between diffusion coordinates of the datasets by considering graph (or manifold) Fourier coefficients of corresponding data features. We then orthogonalize this correlation matrix to form an isometric transformation between the diffusion maps of the datasets. Finally, we apply this transformation to the diffusion coordinates and construct a unified diffusion geometry of the datasets together. We show that this approach successfully corrects misalignment artifacts, and allows for integrated data.

ARM: Augment-REINFORCE-Merge Gradient for Stochastic Binary Networks

No tl;dr =[

To backpropagate the gradients through stochastic binary layers, we propose the augment-REINFORCE-merge (ARM) estimator that is unbiased and has low variance. Exploiting data augmentation, REINFORCE, and reparameterization, the ARM estimator achieves adaptive variance reduction for Monte Carlo integration by merging two expectations via common random numbers. The variance-reduction mechanism of the ARM estimator can also be attributed to antithetic sampling in an augmented space. Experimental results show the ARM estimator provides state-of-the-art performance in auto-encoding variational Bayes and maximum likelihood inference, for discrete latent variable models with one or multiple stochastic binary layers. Python code is available at https://github.com/ABC-anonymous-1.

Context-aware Forecasting for Multivariate Stationary Time-series

tl;dr In order to forecast multivariate stationary time-series we learn embeddings containing contextual features within a RNN; we apply the framework on public transportation data

The domain of time-series forecasting has been extensively studied because it is of fundamental importance in many real-life applications. Weather prediction, traffic flow forecasting or sales are compelling examples of sequential phenomena. Predictive models generally make use of the relations between past and future values. However, in the case of stationary time-series, observed values also drastically depend on a number of exogenous features that can be used to improve forecasting quality. In this work, we propose a change of paradigm which consists in learning such features in embeddings vectors within recurrent neural networks. We apply our framework to forecast smart cards tap-in logs in the Parisian subway network. Results show that context-embedded models perform quantitatively better in one-step ahead and multi-step ahead forecasting.

Opportunistic Learning: Budgeted Cost-Sensitive Learning from Data Streams

tl;dr An online algorithm for cost-aware feature acquisition and prediction

In many real-world learning scenarios, features are only acquirable at a cost constrained under a budget. In this paper, we propose a novel approach for cost-sensitive feature acquisition at the prediction-time. The suggested method acquires features incrementally based on a context-aware feature-value function. We formulate the problem in the reinforcement learning paradigm, and introduce a reward function based on the utility of each feature. Specifically, MC dropout sampling is used to measure expected variations of the model uncertainty which is used as a feature-value function. Furthermore, we suggest sharing representations between the class predictor and value function estimator networks. The suggested approach is completely online and is readily applicable to stream learning setups. The solution is evaluated on three different datasets including the well-known MNIST dataset as a benchmark as well as two cost-sensitive datasets: Yahoo Learning to Rank and a dataset in the medical domain for diabetes classification. According to the results, the proposed method is able to efficiently acquire features and make accurate predictions.

Deep Graph Infomax

tl;dr A new method for unsupervised representation learning on graphs, relying on maximizing mutual information between local and global representations in a graph. State-of-the-art results, competitive with supervised learning.

We present Deep Graph Infomax (DGI), a general approach for learning node representations within graph-structured data in an unsupervised manner. DGI relies on maximizing mutual information between patch representations and corresponding high-level summaries of graphs---both derived using established graph convolutional network architectures. The learnt patch representations summarize subgraphs centered around nodes of interest, and can thus be reused for downstream node-wise learning tasks. In contrast to most prior approaches to graph representation learning, DGI does not rely on random walks, and is readily applicable to both transductive and inductive learning setups. We demonstrate competitive performance on a variety of node classification benchmarks, which at times even exceeds the performance of supervised learning.

tl;dr Using noise to define the divergence between distributions with different support.

For distributions $p$ and $q$ with different support, the divergence $\div{p}{q}$ generally will not exist. We define a spread divergence $\sdiv{p}{q}$ on modified $p$ and $q$ and describe sufficient conditions for the existence of such a divergence. We give examples of using a spread divergence to train implicit generative models, including linear models (Principal Components Analysis and Independent Components Analysis) and non-linear models (Deep Generative Networks).

The Effectiveness of Pre-Trained Code Embeddings

tl;dr Researchers exploring natural language processing techniques applied to source code are not using any form of pre-trained embeddings, we show that they should be.

Word embeddings are widely used in machine learning based natural language processing systems. It is common to use pre-trained word embeddings which provide benefits such as reduced training time and improved overall performance. There has been a recent interest in applying natural language processing techniques to programming languages. However, none of this recent work uses pre-trained embeddings on code tokens. Using an extreme summarization task, we show that using pre-trained embeddings on code tokens provides the same benefits as it does to natural languages, achieving: over 1.8x speedup, 5% relative performance improvement, and resistance to over-fitting. We also show that the choice of programming language used for the embeddings does not have to match that of the task to achieve these benefits.

Ergodic Measure Preserving Flows

tl;dr A novel computational scalable inference framework for training deep generative models and general statistical inference.

Training probabilistic models with neural network components is intractable in most cases and requires to use approximations such as Markov chain Monte Carlo (MCMC), which is not scalable and requires significant hyper-parameter tuning, or mean-field variational inference (VI), which is biased. While there has been attempts at combining both approaches, the resulting methods have some important limitations in theory and in practice. As an alternative, we propose a novel method which is scalable, like mean-field VI, and, due to its theoretical foundation in ergodic theory, is also asymptotically accurate, like MCMC. We test our method on popular benchmark problems with deep generative models and Bayesian neural networks. Our results show that we can outperform existing approximate inference methods.

Understanding Opportunities for Efficiency in Single-image Super Resolution Networks

tl;dr We build an understanding of resource-efficient techniques on Super-Resolution

A successful application of convolutional architectures is to increase the resolution of single low-resolution images -- a vision task called super-resolution (SR). Naturally, SR is of value to resource constrained devices like mobile phones, electronic photograph frames and hoe televisions to enhance image quality. However, SR demands perhaps the most extreme amounts of memory and compute operations of any mainstream vision task known today. And this in-turn prevents SR from being deployed to devices that require them. In this paper, we perform one of the only systematic studies of system resource efficiency for SR, within the context of a variety of architectural and low-precision approaches originally developed for discriminative neural networks. We present a rich set of insights, representative SR architectures and efficiency trade-offs; for example, highly compact SR models suitable for DSPs and FPGAs -- along with SR models suitable for smartphones that are 3x smaller than those of comparable quality found in the literature. Collectively, we believe these results provides the foundation for further research into the little explored area of resource efficiency for SR.

Backprop with Approximate Activations for Memory-efficient Network Training

tl;dr An algorithm to reduce the amount of memory required for training deep networks, based on an approximation strategy.

With innovations in architecture design, deeper and wider neural network models deliver improved performance on a diverse variety of tasks. But the increased memory footprint of these models presents a challenge during training, when all intermediate layer activations need to be stored for back-propagation. Limited GPU memory forces practitioners to make sub-optimal choices: either train inefficiently with smaller batches of examples; or limit the architecture to have lower depth and width, and fewer layers at higher spatial resolutions. This work introduces an approximation strategy that significantly reduces a network's memory footprint during training, but has negligible effect on training performance and computational expense. During the forward pass, we replace activations with lower-precision approximations immediately after they have been used by subsequent layers, thus freeing up memory. The approximate activations are then used during the backward pass. This approach limits the accumulation of errors across the forward and backward pass---because the forward computation across the network still happens at full precision, and the approximation has a limited effect when computing gradients to a layer's input. Experiments, on CIFAR and ImageNet, show that using our approach with 8- and even 4-bit fixed-point approximations of 32-bit floating-point activations has only a minor effect on training and validation performance, while affording significant savings in memory usage.

REVISTING NEGATIVE TRANSFER USING ADVERSARIAL LEARNING

tl;dr We look at negative transfer from a domain adaptation point of view to derive an adversarial learning algorithm.

An unintended consequence of feature sharing is the model fitting to correlated tasks within the dataset, termed negative transfer. In this paper, we revisit the problem of negative transfer in multitask setting and find that its corrosive effects are applicable to a wide range of linear and non-linear models, including neural networks. We first study the effects of negative transfer in a principled way and show that previously proposed counter-measures are insufficient, particularly for trainable features. We propose an adversarial training approach to mitigate the effects of negative transfer by viewing the problem in a domain adaptation setting. Finally, empirical results on attribute prediction multi-task on AWA and CUB datasets further validate the need for correcting negative sharing in an end-to-end manner.

A Main/Subsidiary Network Framework for Simplifying Binary Neural Networks

tl;dr we define the filter-level pruning problem for binary neural networks for the first time and propose method to solve it.

To reduce memory footprint and run-time latency, techniques such as neural net-work pruning and binarization have been explored separately. However, it is un-clear how to combine the best of the two worlds to get extremely small and efficient models. In this paper, we, for the first time, define the filter-level pruning problem for binary neural networks, which cannot be solved by simply migrating existing structural pruning methods for full-precision models. A novel learning-based approach is proposed to prune filters in our main/subsidiary network frame-work, where the main network is responsible for learning representative features to optimize the prediction performance, and the subsidiary component works as a filter selector on the main network. To avoid gradient mismatch when training the subsidiary component, we propose a layer-wise and bottom-up scheme. We also provide the theoretical and experimental comparison between our learning-based and greedy rule-based methods. Finally, we empirically demonstrate the effectiveness of our approach applied on several binary models, including binarizedNIN, VGG-11, and ResNet-18, on various image classification datasets. For bi-nary ResNet-18 on ImageNet, we use 78.6% filters but can achieve slightly better test error 49.87% (50.02%-0.15%) than the original model

Multiple-Attribute Text Rewriting

tl;dr A system for rewriting text conditioned on multiple controllable attributes

The dominant approach to unsupervised "style transfer" in text is based on the idea of learning a latent representation, which is independent of the attributes specifying its "style". In this paper, we show that this condition is not necessary and is not always met in practice, even with domain adversarial training, that explicitly aims at learning such disentangled representations. We thus propose a new model that controls several factors of variation in textual data where this condition on disentanglement is replaced with a simpler mechanism based on back-translation. Our method allows control over multiple attributes, like gender, sentiment, product type, etc., and a more fine-grained control on the trade-off between content preservation and change of style with a pooling operator in the latent space. Our experiments demonstrate that the fully entangled model produces better generations, even when tested on new and more challenging benchmarks comprising reviews with multiple sentences and multiple attributes.

R ESIDUAL NETWORKS CLASSIFY INPUTS BASED ON THEIR NEURAL TRANSIENT DYNAMICS

No tl;dr =[

In this study, we analyze the input-output behavior of residual networks from a dynamical system point of view by disentangling the residual dynamics from the output activities before the classification stage. For a network with simple skip connections between every successive layer, and for logistic activation function, and shared weights between layers, we show analytically that there is a cooperation and competition dynamics between residuals corresponding to each input dimension. Interpreting these kind of networks as nonlinear filters, the steady state value of the residuals in the case of attractor networks are indicative of the common features between different input dimensions that the network has observed during training, and has encoded in those components. In cases where residuals do not converge to an attractor state, their internal dynamics are separable for each input class, and the network can reliably approximate the output. We bring analytical and empirical evidence that residual networks classify inputs based on the integration of the transient dynamics of the residuals. Different inputs are considered as different initial conditions that undergo different transitions through the network, and finally end up in different representations in the output layer. These transitions are critical in assigning the right class to the data. Based on these findings, we also develop a new method to adjust the depth for residual networks during training. As it turns out, after pruning the depth of a Resnet using this algorithm,the network is still capable of classifying inputs with a high accuracy.

Stochastic Optimization of Sorting Networks via Continuous Relaxations

tl;dr We provide a continuous relaxation to the sorting operator, enabling end-to-end, gradient-based stochastic optimization.

Sorting input objects is an important step within many machine learning pipelines. However, the sorting operator is non-differentiable w.r.t. its inputs, which prohibits end-to-end gradient-based optimization. In this work, we propose a general-purpose continuous relaxation of the output of the sorting operator from permutation matrices to the set of "unimodal matrices". Further, we use this relaxation to enable more efficient stochastic optimization over the combinatorially large space of permutations. In particular, we derive a reparameterized gradient estimator for the widely used Plackett-Luce family of distributions. We demonstrate the usefulness of our framework on three tasks that require learning semantic orderings of high-dimensional objects.

Dimension-Free Bounds for Low-Precision Training

tl;dr we proved dimension-independent bounds for low-precision training algorithms

Low-precision training is a promising way of decreasing the time and energy cost of training machine learning models. Previous work has analyzed low-precision training algorithms, such as low-precision stochastic gradient descent, and derived theoretical bounds on their convergence rates. These bounds tend to depend on the dimension of the model $d$ in that the number of bits needed to achieve a particular error bound increases as $d$ increases. This is undesirable because a motivating application for low-precision training is large-scale models, such as deep learning, where $d$ can be huge. In this paper, we prove dimension-independent bounds for low-precision training algorithms that use fixed-point arithmetic, which lets us better understand what affects the convergence of these algorithms as parameters scale. Our methods also generalize naturally to let us prove new convergence bounds on low-precision training with other quantization schemes, such as low-precision floating-point computation and logarithmic quantization.

SENSE: SEMANTICALLY ENHANCED NODE SEQUENCE EMBEDDING

tl;dr Node sequence embedding mechanism that captures both graph and text properties.

Effectively capturing graph node sequences in the form of vector embeddings is critical to many applications. We achieve this by (i) first learning vector embeddings of single graph nodes and (ii) then composing them to compactly represent node sequences. Specifically, we propose SENSE-S (Semantically Enhanced Node Sequence Embedding - for Single nodes), a skip-gram based novel embedding mechanism, for single graph nodes that co-learns graph structure as well as their textual descriptions. We demonstrate that SENSE-S vectors increase the accuracy of multi-label classification tasks by up to 50% and link-prediction tasks by up to 78% under a variety of scenarios using real datasets. Based on SENSE-S, we next propose generic SENSE to compute composite vectors that represent a sequence of nodes, where preserving the node order is important. We prove that this approach is efficient in embedding node sequences, and our experiments on real data confirm its high accuracy in node order decoding.

Latent Domain Transfer: Crossing modalities with Bridging Autoencoders

tl;dr Conditional VAE on top of latent spaces of pre-trained generative models that enables transfer between drastically different domains while preserving locality and semantic alignment.

Domain transfer is a exciting and challenging branch of machine learning because models must learn to smoothly transfer between domains, preserving local variations and capturing many aspects of variation without labels. However, most successful applications to date require the two domains to be closely related (ex. image-to-image, video-video), utilizing similar or shared networks to transform domain specific properties like texture, coloring, and line shapes. Here, we demonstrate that it is possible to transfer across modalities (ex. image-to-audio) by first abstracting the data with latent generative models and then learning transformations between latent spaces. We find that a simple variational autoencoder is able to learn a shared latent space to bridge between two generative models in an unsupervised fashion, and even between different types of models (ex. variational autoencoder and a generative adversarial network). We can further impose desired semantic alignment of attributes with a linear classifier in the shared latent space. The proposed variation autoencoder enables preserving both locality and semantic alignment through the transfer process, as shown in the qualitative and quantitative evaluations. Finally, the hierarchical structure decouples the cost of training the base generative models and semantic alignments, enabling computationally efficient and data efficient retraining of personalized mapping functions.

A fast quasi-Newton-type method for large-scale stochastic optimisation

No tl;dr =[

During recent years there has been an increased interest in stochastic adaptations of limited memory quasi-Newton methods, which compared to pure gradient-based routines can improve the convergence by incorporating second order information. In this work we propose a direct least-squares approach conceptually similar to the limited memory quasi-Newton methods, but that computes the search direction in a slightly different way. This is achieved in a fast and numerically robust manner by maintaining a Cholesky factor of low dimension. This is combined with a stochastic line search relying upon fulfilment of the Wolfe condition in a backtracking manner, where the step length is adaptively modified with respect to the optimisation progress. We support our new algorithm by providing several theoretical results guaranteeing its performance. The performance is demonstrated on real-world benchmark problems which shows improved results in comparison with already established methods.

An Efficient and Margin-Approaching Zero-Confidence Adversarial Attack

tl;dr This paper introduces MarginAttack, a stronger and faster zero-confidence adversarial attack.

There are two major paradigms of white-box adversarial attacks that attempt to impose input perturbations. The first paradigm, called the fix-perturbation attack, crafts adversarial samples within a given perturbation level. The second paradigm, called the zero-confidence attack, finds the smallest perturbation needed to cause misclassification, also known as the margin of an input feature. While the former paradigm is well-resolved, the latter is not. Existing zero-confidence attacks either introduce significant approximation errors, or are too time-consuming. We therefore propose MarginAttack, a zero-confidence attack framework that is able to compute the margin with improved accuracy and efficiency. Our experiments show that MarginAttack is able to compute a smaller margin than the state-of-the-art zero-confidence attacks, and matches the state-of-the-art fix-perturbation attacks. In addition, it runs significantly faster than the Carlini-Wagner attack, currently the most accurate zero-confidence attack algorithm.

Importance Resampling for Off-policy Policy Evaluation

tl;dr A resampling approach for off-policy policy evaluation in reinforcement learning.

Importance sampling is a common approach to off-policy learning in reinforcement learning. While it is consistent and unbiased, it can result in high variance updates to the parameters for the value function. Weighted importance sampling (WIS) has been explored to reduce variance for off-policy policy evaluation, but only for linear value function approximation. In this work, we explore a resampling strategy to reduce variance, rather than a reweighting strategy. We propose Importance Resampling (IR) for off-policy learning, that resamples experience from the replay buffer and applies a standard on-policy update. The approach avoids using importance sampling ratios directly in the update, instead correcting the distribution over transitions before the update. We characterize the bias and consistency of the our estimator, particularly compared to WIS. We then demonstrate in several toy domains that IR has improved sample efficiency and parameter sensitivity, as compared to several baseline WIS estimators and to IS. We conclude with a demonstration showing IR improves over IS for learning a value function from images in a racing car simulator.

Mode Normalization

tl;dr We present a novel normalization method for deep neural networks that is robust to multi-modalities in intermediate feature distributions.

Normalization methods are a central building block in the deep learning toolbox. They accelerate and stabilize training, while decreasing the dependence on manually tuned learning rate schedules. When learning from multi-modal distributions, the effectiveness of batch normalization (BN), arguably the most prominent normalization method, is reduced. As a remedy, we propose a more flexible approach: by extending the normalization to more than a single mean and variance, we detect modes of data on-the-fly, jointly normalizing samples that share common features. We demonstrate that our method outperforms BN and other widely used normalization techniques in several experiments, including single and multi-task datasets.

A Variational Inequality Perspective on Generative Adversarial Networks

tl;dr We cast GANs in the variational inequality framework and import techniques from this literature to optimize GANs better; we give algorithmic extensions and empirically test their performance for training GANs.

Generative adversarial networks (GANs) form a generative modeling approach known for producing appealing samples, but they are notably difficult to train. One common way to tackle this issue has been to propose new formulations of the GAN objective. Yet, surprisingly few studies have looked at optimization methods designed for this adversarial training. In this work, we cast GAN optimization problems in the general variational inequality framework. Tapping into the mathematical programming literature, we counter some common misconceptions about the difficulties of saddle point optimization and propose to extend methods designed for variational inequalities to the training of GANs. We apply averaging, extrapolation and a novel computationally cheaper variant that we call extrapolation from the past to the stochastic gradient method (SGD) and Adam.

Soft Q-Learning with Mutual-Information Regularization

No tl;dr =[

We propose a reinforcement learning (RL) algorithm that uses mutual-information regularization to optimize the prior action distribution for better performance and exploration. Entropy-based regularization has previously been shown to improve both exploration and robustness in challenging sequential decision-making tasks. It does so by encouraging policies to put probability mass on all actions. However, entropy regularization might be undesirable when actions have significantly different importance. In this paper, we propose a theoretically motivated framework that dynamically weights the importance of actions by using the mutual-information. In particular, we express the RL problem as an inference problem where the prior probability distribution over actions is subject to optimization. We show that the prior optimization introduces a mutual-information regularizer in the RL objective. This regularizer encourages the policy to be close to a non-uniform distribution that assigns higher probability mass to more important actions. We empirically demonstrate that our method significantly improves over entropy regularization methods, attaining state-of-the-art performance.

INVASE: Instance-wise Variable Selection using Neural Networks

No tl;dr =[

The advent of big data brings with it data with more and more dimensions and thus a growing need to be able to efficiently select which features to use for a variety of problems. While global feature selection has been a well-studied problem for quite some time, only recently has the paradigm of instance-wise feature selection been developed. In this paper, we propose a new instance-wise feature selection method, which we term INVASE. INVASE consists of 3 neural networks, a selector network, a predictor network and a baseline network which are used to train the selector network using the actor-critic methodology. Using this methodology, INVASE is capable of flexibly discovering feature subsets of a different size for each instance, which is a key limitation of existing state-of-the-art methods. We demonstrate through a mixture of synthetic and real data experiments that INVASE significantly outperforms state-of-the-art benchmarks.

Model-Predictive Policy Learning with Uncertainty Regularization for Driving in Dense Traffic

No tl;dr =[

Learning a policy using only observational data is challenging because the distribution of states it induces at execution time may differ from the distribution observed during training. In this work, we propose to train a policy while explicitly penalizing the mismatch between these two distributions over a fixed time horizon. We do this by using a learned model of the environment dynamics which is unrolled for multiple time steps, and training a policy network to minimize a differentiable cost over this rolled-out trajectory. This cost contains two terms: a policy cost which represents the objective the policy seeks to optimize, and an uncertainty cost which represents its divergence from the states it is trained on. We propose to measure this second cost by using the uncertainty of the dynamics model about its own predictions, using recent ideas from uncertainty estimation for deep networks. We evaluate our approach using a large-scale observational dataset of driving behavior recorded from traffic cameras, and show that we are able to learn effective driving policies from purely observational data, with no environment interaction.

Object detection deep learning networks for Optical Character Recognition

tl;dr Yolo / RCNN neural network for object detection adapted to the task of OCR

In this article, we show how we applied a simple approach coming from deep learning networks for object detection to the task of optical character recognition in order to build image features taylored for documents. In contrast to scene text reading in natural images using networks pretrained on ImageNet, our document reading is performed with small networks inspired by MNIST digit recognition challenge, at a small computational budget and a small stride. The object detection modern frameworks allow a direct end-to-end training, with no other algorithm than the deep learning and the non-max-suppression algorithm to filter the duplicate predictions. The trained weights can be used for higher level models, such as, for example, document classification, or document segmentation.

Deep Neuroevolution: Genetic Algorithms are a Competitive Alternative for Training Deep Neural Networks for Reinforcement Learning

No tl;dr =[

Deep artificial neural networks (DNNs) are typically trained via gradient-based learning algorithms, namely backpropagation. Evolution strategies (ES) can rival backprop-based algorithms such as Q-learning and policy gradients on challenging deep reinforcement learning (RL) problems. However, ES can be considered a gradient-based algorithm because it performs stochastic gradient descent via an operation similar to a finite-difference approximation of the gradient. That raises the question of whether non-gradient-based evolutionary algorithms can work at DNN scales. Here we demonstrate they can: we evolve the weights of a DNN with a simple, gradient-free, population-based genetic algorithm (GA) and it performs well on hard deep RL problems, including Atari and humanoid locomotion. The Deep GA successfully evolves networks with over four million free parameters, the largest neural networks ever evolved with a traditional evolutionary algorithm. These results (1) expand our sense of the scale at which GAs can operate, (2) suggest intriguingly that in some cases following the gradient is not the best choice for optimizing performance, and (3) make immediately available the multitude of neuroevolution techniques that improve performance. We demonstrate the latter by showing that combining DNNs with novelty search, which encourages exploration on tasks with deceptive or sparse reward functions, can solve a high-dimensional problem on which reward-maximizing algorithms (e.g.\ DQN, A3C, ES, and the GA) fail. Additionally, the Deep GA is faster than ES, A3C, and DQN (it can train Atari in {\raise.17ex\hbox{$\scriptstyle\sim$}}4 hours on one workstation or {\raise.17ex\hbox{$\scriptstyle\sim$}}1 hour distributed on 720 cores), and enables a state-of-the-art, up to 10,000-fold compact encoding technique.

EDDI: Efficient Dynamic Discovery of High-Value Information with Partial VAE

No tl;dr =[

Making decisions requires information relevant to the task at hand. Many real-life decision-making situations allow acquiring further relevant information at a specific cost. For example, in assessing the health status of a patient we may decide to take additional measurements such as diagnostic tests or imaging scans before making a final assessment. More information that is relevant allows for better decisions but it may be costly to acquire all of this information. How can we trade off the desire to make good decisions with the option to acquire further information at a cost? To this end, we propose a principled framework, named EDDI (Efficient Dynamic Discovery of high-value Information), based on the theory of Bayesian experimental design. In EDDI we propose a novel partial variational autoencoder (Partial VAE), to efficiently handle missing data over varying subsets of known information. EDDI combines this Partial VAE with an acquisition function that maximizes expected information gain on a set of target variables. EDDI is efficient and demonstrates that dynamic discovery of high-value information is possible; we show cost reduction at the same decision quality and improved decision quality at the same cost in benchmarks and in two health-care applications.. We believe there is great potential for realizing these gains in real-world decision support systems.

Unsupervised Latent Tree Induction with Deep Inside-Outside Recursive Auto-Encoders

tl;dr In this work we propose deep inside-outside recursive auto-encoders(DIORA) a fully unsupervised method of discovering syntax while simultaneously learning representations for discovered constituents.

Syntax is a powerful abstraction for language understanding. Many downstream tasks require segmenting input text into meaningful constituent chunks (e.g., noun phrases or entities); more generally, models for learning semantic representations of text benefit from integrating syntax in the form of parse trees (e.g., tree-LSTMs). Supervised parsers have traditionally been used to obtain these trees, but lately interest has increased in unsupervised methods that induce syntactic representations directly from unlabeled text. To this end, we propose the deep inside-outside recursive autoencoder (DIORA), a fully-unsupervised method for discovering syntax that simultaneously learns representations for constituents within the induced tree. Unlike many prior approaches, DIORA does not rely on supervision from auxiliary downstream tasks and is thus not constrained to particular domains. Furthermore, competing approaches do not learn explicit phrase representations along with tree structures, which limits their applicability to phrase-based tasks. Extensive experiments on unsupervised parsing, segmentation, and phrase clustering demonstrate the efficacy of our method. DIORA achieves the state of the art in unsupervised parsing (46.9 F1) on the benchmark WSJ dataset.

Learning Programmatically Structured Representations with Perceptor Gradients

No tl;dr =[

We present the perceptor gradients algorithm -- a novel approach to learning symbolic representations based on the idea of decomposing an agent's policy into i) a perceptor network extracting symbols from raw observation data and ii) a task encoding program which maps the input symbols to output actions. We show that the proposed algorithm is able to learn representations that can be directly fed into a Linear-Quadratic Regulator (LQR) or a general purpose A* planner. Our experimental results confirm that the perceptor gradients algorithm is able to efficiently learn transferable symbolic representations as well as generate new observations according to a semantically meaningful specification.

Interpreting Adversarial Robustness: A View from Decision Surface in Input Space

No tl;dr =[

One popular hypothesis of neural network generalization is that the flat local minima of loss surface in parameter space leads to good generalization. However, we demonstrate that loss surface in parameter space has no obvious relationship with generalization, especially under adversarial settings. Through visualizing decision surfaces in both parameter space and input space, we instead show that the geometry property of decision surface in input space correlates well with the adversarial robustness. We then propose an adversarial robustness indicator, which can evaluate a neural network's intrinsic robustness property without testing its accuracy under adversarial attacks. Guided by it, we further propose our robust training method. Without involving adversarial training, our method could enhance network's intrinsic adversarial robustness against various adversarial attacks.

Encoding Category Trees Into Word-Embeddings Using Geometric Approach

tl;dr we show a geometric method to perfectly encode categroy tree information into pre-trained word-embeddings.

We present a novel method to implicitly encode a tree-structured category information into word-embeddings, resulting in super-dimensional ball representations ($n$-ball embedding for short). Inclusion relations among $n$-balls precisely encode subordinate relations among categories. The cosine similarity function is enriched by category information. A large $n$-ball dataset is constructed using geometrical method, which achieves zero energy cost in embedding tree structures into word embedding. A new benchmark dataset is created for predicting the category of unknown words. Experiments show that $n$-ball embeddings, carried with category information, significantly out-perform word-embeddings in the neighbourhood test, while only slightly change the original word-embeddings. Experiment results also show that $n$-ball embeddings demonstrate surprisingly good performance in validating the category of unknown word. Source codes and data-sets are free for public access \url{https://github.com/gnodisnait/nball4tree.git} and \url{https://github.com/gnodisnait/bp94nball.git}.

Unsupervised Neural Multi-Document Abstractive Summarization

tl;dr We propose an end-to-end neural model for unsupervised multi-document abstractive summarization, applying it to business and product reviews.

Abstractive summarization has been studied using neural sequence transduction methods with datasets of large, paired document-summary examples. However, such datasets are rare and the models trained from them do not generalize to other domains. Recently, some progress has been made in learning sequence-to-sequence mappings with only unpaired examples. In our work, we consider the setting where there are only documents and no summaries provided and propose an end-to-end, neural model architecture to perform unsupervised abstractive summarization. Our proposed model consists of an auto-encoder trained so that the mean of the representations of the input documents decodes to a reasonable summary. We consider variants of the proposed architecture and perform an ablation study to show the importance of specific components. We apply our model to the summarization of business and product reviews and show that the generated summaries are fluent, show relevancy in terms of word-overlap, representative of the average sentiment of the input documents, and are highly abstractive compared to baselines. The code to reproduce results is available at github.com/REDACTED.

FEATURE PRIORITIZATION AND REGULARIZATION IMPROVE STANDARD ACCURACY AND ADVERSARIAL ROBUSTNESS

tl;dr We propose a model that employs feature prioritization and regularization to improve the adversarial robustness and the standard accuracy.

Adversarial training has been successfully applied to build robust models at a certain cost. While the robustness of a model increases, the standard classification accuracy declines. This phenomenon is suggested to be an inherent trade-off between standard accuracy and robustness. We propose a model that employs feature prioritization by a nonlinear attention module and L2 regularization as implicit denoising to improve the adversarial robustness and the standard accuracy relative to adversarial training. Focusing sharply on the regions of interest, the attention maps encourage the model to rely heavily on features extracted from the most relevant areas while suppressing the unrelated background. Penalized by a regularizer, the model extracts similar features for the natural and adversarial images, effectively ignoring the added perturbation. In addition to qualitative evaluation, we also propose a novel experimental strategy that quantitatively demonstrates that our model is almost ideally aligned with salient data characteristics. Additional experimental results illustrate the power of our model relative to the state of the art methods

Metric-Optimized Example Weights

No tl;dr =[

Real-world machine learning applications often have complex test metrics, and may have training and test data that follow different distributions. We propose addressing these issues by using a weighted loss function with a standard convex loss, but with weights on the training examples that are learned to optimize the test metric of interest on the validation set. These metric-optimized example weights can be learned for any test metric, including black box losses and customized metrics for specific applications. We illustrate the performance of our proposal with public benchmark datasets and real-world applications with domain shift and custom loss functions that balance multiple objectives, impose fairness policies, and are non-convex and non-decomposable.

Expressiveness in Deep Reinforcement Learning

No tl;dr =[

Representation learning in reinforcement learning (RL) algorithms focuses on extracting useful features for choosing good actions. Expressive representations are essential for learning well-performed policies. In this paper, we study the relationship between the state representation assigned by the state extractor and the performance of the RL agent. We observe that representations assigned by the better state extractor are more scattered than which assigned by the worse one. Moreover, RL agents achieving high performances always have high rank matrices which are composed by their representations. Based on our observations, we formally define expressiveness of the state extractor as the rank of the matrix composed by representations. Therefore, we propose to promote expressiveness so as to improve algorithm performances, and we call it Expressiveness Promoted DRL. We apply our method on both policy gradient and value-based algorithms, and experimental results on 55 Atari games show the superiority of our proposed method.

A Unified View of Deep Metric Learning via Gradient Analysis

No tl;dr =[

Loss functions play a pivotal role in deep metric learning (DML). A large variety of loss functions have been proposed in DML recently. However, it remains difficult to answer this question: what are the intrinsic differences among these loss functions?This paper answers this question by proposing a unified perspective to rethink deep metric loss functions. We show theoretically that most DML methods in deep metric learning, in view of gradient equivalence, are essentially weight assignment strategies of training pairs. Based on this unified view, we revisit several typical DML methods and disclose their hidden drawbacks. Moreover, we point out the key components of an effective DML approach which drives us to propose our weight assignment framework. We evaluate our method on image retrieval tasks, and show that it outperforms the state-of-the-art DML approaches by a significant margin on the CUB-200-2011, Cars-196, Stanford Online Products and In-Shop Clothes Retrieval datasets.

Diffusion Scattering Transforms on Graphs

tl;dr Stability of scattering transform representations of graph data to deformations of the underlying graph support.

Stability is a key aspect of data analysis. In many applications, the natural notion of stability is geometric, as illustrated for example in computer vision. Scattering transforms construct deep convolutional representations which are certified stable to input deformations. This stability to deformations can be interpreted as stability with respect to changes in the metric structure of the domain. In this work, we show that scattering transforms can be generalized to non-Euclidean domains using diffusion wavelets, while preserving a notion of stability with respect to metric changes in the domain, measured with diffusion maps. The resulting representation is stable to metric perturbations of the domain while being able to capture ''high-frequency'' information, akin to the Euclidean Scattering.

tl;dr MAML is great, but it has many problems, we solve many of those problems and as a result we learn most hyper parameters end to end, speed-up training and inference and set a new SOTA in few-shot learning

The field of few-shot learning has recently seen substantial advancements. Most of these advancements came from casting few-shot learning as a meta-learning problem.Model Agnostic Meta Learning or MAML is currently one of the best approaches for few-shot learning via meta-learning. MAML is simple, elegant and very powerful, however, it has a variety of issues, such as being very sensitive to neural network architectures, often leading to instability during training, requiring arduous hyperparameter searches to stabilize training and achieve high generalization and being very computationally expensive at both training and inference times. In this paper, we propose various modifications to MAML that not only stabilize the system, but also substantially improve the generalization performance, convergence speed and computational overhead of MAML, which we call MAML++.

tl;dr We propose a new attribution method that removes noise from saliency maps through layer-wise thresholding during backpropagation.

Saliency map, or the gradient of the score function with respect to the input, is the most basic means of interpreting deep neural network decisions. However, saliency maps are often visually noisy. Although several hypotheses were proposed to account for this phenomenon, there is no work that provides a rigorous analysis of noisy saliency maps. This may be a problem as numerous advanced attribution methods were proposed under the assumption that the existing hypotheses are true. In this paper, we identify the cause of noisy saliency maps. Then, we propose Rectified Gradient, a simple method that significantly improves saliency maps by alleviating that cause. Experiments showed effectiveness of our method and its superiority to other attribution methods. Codes and examples for the experiments will be released in public.

RedSync : Reducing Synchronization Traffic for Distributed Deep Learning

tl;dr We proposed an implementation to accelerate DNN data parallel training by reducing communication bandwidth requirement.

Data parallelism has become a dominant method to scale Deep Neural Network (DNN) training across multiple nodes. Since the synchronization of the local models or gradients can be a bottleneck for large-scale distributed training, compressing communication traffic has gained widespread attention recently. Among several recent proposed compression algorithms, Residual Gradient Compression (RGC) is one of the most successful approaches---it can significantly compress the transmitting message size (0.1% of the gradient size) of each node and still preserve accuracy. However, the literature on compressing deep networks focuses almost exclusively on achieving good compression rate, while the efficiency of RGC in real implementation has been less investigated. In this paper, we develop an RGC method that achieves significant training time improvement in real-world multi-GPU systems. Our proposed RGC system design called RedSync, introduces a set of optimizations to reduce communication bandwidth while introducing limited overhead. We examine the performance of RedSync on two different multiple GPU platforms, including a supercomputer and a multi-card server. Our test cases include image classification on Cifar10 and ImageNet, and language modeling tasks on Penn Treebank and Wiki2 datasets. For DNNs featured with high communication to computation ratio, which has long been considered with poor scalability, RedSync shows significant performance improvement.

tl;dr A new nonlinear defense against adversarial attacks.

The linear and non-flexible nature of deep convolutional models makes them vulnerable to carefully crafted adversarial perturbations. To tackle this problem, in this paper, we propose a nonlinear radial basis convolutional feature transformation by learning the Mahalanobis distance function that maps the input convolutional features from the same class into tight clusters. In such a space, the clusters become compact and well-separated, which prevent small adversarial perturbations from forcing a sample to cross the decision boundary. We test the proposed method on three publicly available image classification and segmentation data-sets namely, MNIST, ISBI ISIC skin lesion, and NIH ChestX-ray14. We evaluate the robustness of our method to different gradient (targeted and untargeted) and non-gradient based attacks and compare it to several non-gradient masking defense strategies. Our results demonstrate that the proposed method can boost the performance of deep convolutional neural networks against adversarial perturbations without accuracy drop on clean data.

Fixing Posterior Collapse with delta-VAEs

tl;dr Avoid posterior collapse by lower bounding the rate.

Due to the phenomenon of “posterior collapse”, current latent variable generativemodels pose a challenging design choice which trades-off optimizing the ELBObut handicapping the decoders’ capacity and expressivity, or changing the loss tosomething that is not directly minimizing the description length. In this paper wepropose an alternative that utilizes the best, most powerful generative models asdecoders, whilst optimizing the proper variational lower bound all while ensuringthat the latent variables preserve and encode useful information. delta-VAEs pro-posed here achieve this by constraining the variational family for the posterior tohave a minimum distance to the prior, which resembles the classic representationlearning approach of slow feature analysis. We demonstrate the efficacy of our ap-proach at modeling images: learning representations, improving sample quality,and improving state of the art log-likelihood on CIFAR-10 and ImageNet32×32.

Cramer-Wold AutoEncoder

tl;dr Inspired by prior work on Sliced-Wasserstein Autoencoders (SWAE) and kernel smoothing we construct a new generative model – Cramer-Wold AutoEncoder (CWAE).

Assessing distance betweeen the true and the sample distribution is a key component of many state of the art generative models, such as Wasserstein Autoencoder (WAE). Inspired by prior work on Sliced-Wasserstein Autoencoders (SWAE) and kernel smoothing we construct a new generative model – Cramer-Wold AutoEncoder (CWAE). CWAE cost function, based on introduced Cramer-Wold distance between samples, has a simple closed-form in the case of normal prior. As a consequence, while simplifying the optimization procedure (no need of sampling necessary to evaluate the distance function in the training loop), CWAE performance matches quantitatively and qualitatively that of WAE-MMD (WAE using maximum mean discrepancy based distance function) and often improves upon SWAE.

Data Interpretation and Reasoning Over Scientific Plots

tl;dr We created a new dataset for data interpretation over plots and also propose a baseline for the same.

Data Interpretation is an important part of Quantitative Aptitude exams and requires an individual to answer questions grounded in plots such as bar charts, line graphs, scatter plots, \textit{etc}. Recently, there has been an increasing interest in building models which can perform this task by learning from datasets containing triplets of the form \{plot, question, answer\}. Two such datasets have been proposed in the recent past which contain plots generated from synthetic data with limited (i) $x-y$ axes variables (ii) question templates and (iii) answer vocabulary and hence do not adequately capture the challenges posed by this task. To overcome these limitations of existing datasets, we introduce a new dataset containing $9.7$ million question-answer pairs grounded over $270,000$ plots with three main differentiators. First, the plots in our dataset contain a wide variety of realistic $x$-$y$ variables such as CO2 emission, fertility rate, \textit{etc.} extracted from real word data sources such as World Bank, government sites, \textit{etc}. Second, the questions in our dataset are more complex as they are based on templates extracted from interesting questions asked by a crowd of workers using a fraction of these plots. Lastly, the answers in our dataset are not restricted to a small vocabulary and a large fraction of the answers seen at test time are not present in the training vocabulary. As a result, existing models for Visual Question Answering which largely use end-to-end models in a multi-class classification framework cannot be used for this task. We establish initial results on this dataset and emphasize the complexity of the task using a multi-staged modular pipeline with various sub-components to (i) extract relevant data from the plot and convert it to a semi-structured table (ii) combine the question with this table and use compositional semantic parsing to arrive at a logical form from which the answer can be derived. We believe that such a modular framework is the best way to go forward as it would enable the research community to independently make progress on all the sub-tasks involved in plot question answering.

Lipschitz regularized Deep Neural Networks converge and generalize

tl;dr We prove generalization of DNNs by adding a Lipschitz regularization term to the training loss

Generalization of deep neural networks (DNNs) is an open problem which, if solved, could impact the reliability and verification of deep neural network architectures. In this paper, we show that if the usual fidelity term used in training DNNs is augmented by a Lipschitz regularization term, then the networks converge and generalize. The convergence is in the limit as the number of data points, $n\to \infty$, while also allowing the network to grow as needed to fit the data. Two regimes are identified: in the case of clean labels, we prove convergence to the label function which corresponds to zero loss, in the case of corrupted labels which we prove convergence to a regularized label function which is the solution of a limiting variational problem. In both cases, a convergence rate is also provided.

Efficient Dictionary Learning with Gradient Descent

tl;dr We provide an efficient convergence rate for gradient descent on the complete orthogonal dictionary learning objective based on a geometric analysis.

Randomly initialized first-order optimization algorithms are the method of choice for solving many high-dimensional nonconvex problems in machine learning, yet general theoretical guarantees cannot rule out convergence to critical points of poor objective value. For some highly structured nonconvex problems however, the success of gradient descent can be understood by studying the geometry of the objective. We study one such problem -- complete orthogonal dictionary learning, and provide converge guarantees for randomly initialized gradient descent to the neighborhood of a global optimum. The resulting rates scale as low order polynomials in the dimension even though the objective possesses an exponential number of saddle points. This efficient convergence can be viewed as a consequence of negative curvature normal to the stable manifolds associated with saddle points, and we provide evidence that this feature is shared by other nonconvex problems of importance as well.

Link Prediction in Hypergraphs using Graph Convolutional Networks

tl;dr We propose Neural Hyperlink Predictor (NHP). NHP adapts graph convolutional networks for link prediction in hypergraphs

Link prediction in simple graphs is a fundamental problem in which new links between nodes are predicted based on the observed structure of the graph. However, in many real-world applications, there is a need to model relationships among nodes which go beyond pairwise associations. For example, in a chemical reaction, relationship among the reactants and products is inherently higher-order. Additionally, there is need to represent the direction from reactants to products. Hypergraphs provide a natural way to represent such complex higher-order relationships. Even though Graph Convolutional Networks (GCN) have recently emerged as a powerful deep learning-based approach for link prediction over simple graphs, their suitability for link prediction in hypergraphs is unexplored -- we fill this gap in this paper and propose Neural Hyperlink Predictor (NHP). NHP adapts GCNs for link prediction in hypergraphs. We propose two variants of NHP --NHP-U and NHP-D -- for link prediction over undirected and directed hypergraphs, respectively. To the best of our knowledge, NHP-D is the first method for link prediction over directed hypergraphs. Through extensive experiments on multiple real-world datasets, we show NHP's effectiveness.

Accelerated Sparse Recovery Under Structured Measurements

No tl;dr =[

Extensive work on compressed sensing has yielded a rich collection of sparse recovery algorithms, each making different tradeoffs between recovery condition and computational efficiency. In this paper, we propose a unified framework for accelerating various existing sparse recovery algorithms without sacrificing recovery guarantees by exploiting structure in the measurement matrix. Unlike fast algorithms that are specific to particular choices of measurement matrices where the columns are Fourier or wavelet filters for example, the proposed approach works on a broad range of measurement matrices that satisfy a particular property. We precisely characterize this property, which quantifies how easy it is to accelerate sparse recovery for the measurement matrix in question. We also derive the time complexity of the accelerated algorithm, which is sublinear in the signal length in each iteration. Moreover, we present experimental results on real world data that demonstrate the effectiveness of the proposed approach in practice.

Neural Causal Discovery with Learnable Input Noise

No tl;dr =[

Learning causal relations from observational time series with nonlinear interactions and complex causal structures is a key component of human intelligence, and has a wide range of applications. Although neural nets have demonstrated their effectiveness in a variety of fields, their application in learning causal relations has been scarce. This is due to both a lack of theoretical results connecting risk minimization and causality (enabling function approximators like neural nets to apply), and a lack of scalability in prior causal measures to allow for expressive function approximators like neural nets to apply. In this work, we propose a novel causal measure and algorithm using risk minimization to infer causal relations from time series. We prove that under certain conditions, the positiveness of our measure can deduce a stringent direct structural causality. We demonstrate the effectiveness and scalablility of our algorithms to learn nonlinear causal models in synthetic datasets as comparing to other methods, and its effectiveness in inferring causal relations in a video game environment and real-world heart-rate vs. breath-rate and rat brain EEG datasets.

Learning To Plan

No tl;dr =[

We introduce Learning To Plan (L2P), a novel architecture for deep reinforcement learning, that combines model-based and model-free aspects for online planning. Our architecture learns to dynamically construct plans using a learned state-transition model by selecting and traversing between simulated states and actions to maximize valuable information before acting. In contrast to model-free methods, model-based planning lets the agent efficiently test action hypotheses without performing costly trial-and-error in the environment. L2P learns to efficiently form plans by expanding a single action-conditional state transition at a time instead of exhaustively evaluating each action, reducing the required number of state-transitions during planning by up to 96%. We observe various emergent planning patterns used to solve environments, including classical search methods such as breadth-first and depth-first search. Learning To Plan shows improved data efficiency, performance, and generalization to new and unseen domains in comparison to several baselines.

Countering Language Drift via Grounding

tl;dr Grounding helps avoid language drift during fine-tuning natural language agents with policy gradients.

While reinforcement learning (RL) shows a lot of promise for natural language processing—e.g. when fine-tuning natural language systems for optimizing a certain objective—there has been little investigation into potential language drift: when an external reward is used to train a system, the agents’ communication protocol may easily and radically diverge from natural language. By re-casting translation as a communication game, we show that language drift indeed happens when pre-trained agents are fine-tuned with policy gradient methods. We contend that simply adding a "naturalness" constraint to the reward, e.g. by using language model log likelihood, does not fully address the issue, and argue that (perceptual) grounding is required. That is, while language model constraints impose syntactic conformity, they do not lead to semantic correspondence. Our experiments show that grounded models give the best communication performance, while retaining English syntax along with the ability to convey the intended semantics.

RNNs with Private and Shared Representations for Semi-Supervised Sequence Learning

tl;dr This paper focuses upon a traditionally overlooked mechanism -- an architecture with explicitly designed private and shared hidden units designed to mitigate the detrimental influence of the auxiliary unsupervised loss over the main supervised task.

Training recurrent neural networks (RNNs) on long sequences using backpropagation through time (BPTT) remains a fundamental challenge. It has been shown that adding a local unsupervised loss term into the optimization objective makes the training of RNNs on long sequences more effective. While the importance of an unsupervised task can in principle be controlled by a coefficient in the objective function, the gradients with respect to the unsupervised loss term still influence all the hidden state dimensions, which might cause important information about the supervised task to be degraded or erased. Compared to existing semi-supervised sequence learning methods, this paper focuses upon a traditionally overlooked mechanism -- an architecture with explicitly designed private and shared hidden units designed to mitigate the detrimental influence of the auxiliary unsupervised loss over the main supervised task. We achieve this by dividing RNN hidden space into a private space for the supervised task and a shared space for both the supervised and unsupervised tasks. We present extensive experiments with the proposed framework on several long sequence modeling benchmark datasets. Results indicate that the proposed framework can yield performance gains in RNN models where long term dependencies are notoriously challenging to deal with.

Large-Scale Visual Speech Recognition

No tl;dr =[

This work presents a scalable solution to open-vocabulary visual speech recognition. To achieve this, we constructed the largest existing visual speech recognition dataset, consisting of pairs of text and video clips of faces speaking (3,886 hours of video). In tandem, we designed and trained an integrated lipreading system, consisting of a video processing pipeline that maps raw video to stable videos of lips and sequences of phonemes, a scalable deep neural network that maps the lip videos to sequences of phoneme distributions, and a production-level speech decoder that outputs sequences of words. The proposed system achieves a word error rate (WER) of 40.9% as measured on a held-out set. In comparison, professional lipreaders achieve either 86.4% or 92.9% WER on the same dataset when having access to additional types of contextual information. Our approach significantly improves on other lipreading approaches, including variants of LipNet and of Watch, Attend, and Spell (WAS), which are only capable of 89.8% and 76.8% WER respectively.

Feature quantization for parsimonious and meaningful predictive models

tl;dr We tackle discretization of continuous features and grouping of factor levels as a representation learning problem and provide a rigorous way of estimating the best quantization to yield good performance and interpretability.

For regulatory and interpretability reasons, the logistic regression is still widely used by financial institutions to learn the refunding probability of a loan given the applicant's characteristics from historical data. Although logistic regression handles naturally both continuous and categorical data, a preprocessing step to quantize them is usually performed for improving simultaneously prediction accuracy and user interpretability: continuous features are discretized by assigning factor levels to intervals; some levels of categorical features (with numerous levels) are grouped. However, a better predictive accuracy can be reached by embedding this quantization estimation step directly into the predictive estimation step itself. A related information criterion has then to be optimized on a huge and untractable discontinuous quantization set, requiring to introduce a specific two-step optimization strategy: first, the optimization problem is relaxed in order to deal with smooth functions; second, a particular neural network is involved through a stochastic gradient algorithm to optimize the resulting criterion, giving access to good candidates for the initial optimization problem. The good performances of this approach are illustrated on simulated and real data from Crédit Agricole Consumer Finance (a major European historic player in the consumer credit market).

SnapQuant: A Probabilistic and Nested Parameterization for Binary Networks

tl;dr We propose SnapQuant, a reinforcement learning method for training binary weight networks from scratch under the Bayesian deep learning perspective, which approximates the posterior distribution of binary weights instead of a single point estimation.

In this paper, we study the problem of training real binary weight networks (without layer-wise or filter-wise scaling factors) from scratch under the Bayesian deep learning perspective, meaning that the final objective is to approximate the posterior distribution of binary weights rather than reach a point estimation. The proposed method, named as SnapQuant, has two intriguing features: (1) The posterior distribution is parameterized as a policy network trained with a reinforcement learning scheme. During the training phase, we generate binary weights on-the-fly since what we actually maintain is the policy network, and all the binary weights are used in a burn-after-reading style. At the testing phase, we can sample binary weight instances for a given recognition architecture from the learnt policy network. (2) The policy network, which has a nested parameter structure consisting of layer-wise, filter-wise and kernel-wise parameter sharing designs, is applicable to any neural network architecture. Such a nested parameterization explicitly and hierarchically models the joint posterior distribution of binary weights. The performance of SnapQuant is evaluated with several visual recognition tasks including ImageNet. The code will be made publicly available.

Individualized Controlled Continuous Communication Model for Multiagent Cooperative and Competitive Tasks

tl;dr We introduce IC3Net, a single network which can be used to train agents in cooperative, competitive and mixed scenarios. We also show that agents can learn when to communicate using our model.

Learning when to communicate and doing that effectively is essential in multi-agent tasks. Recent works show that continuous communication allows efficient training with back-propagation in multi-agent scenarios, but have been restricted to fully-cooperative tasks. In this paper, we present Individualized Controlled Continuous Communication Model (IC3Net) which has better training efficiency than simple continuous communication model, and can be applied to semi-cooperative and competitive settings along with the cooperative settings. IC3Net controls continuous communication with a gating mechanism and uses individualized rewards foreach agent to gain better performance and scalability while fixing credit assignment issues. Using variety of tasks including StarCraft BroodWars explore and combat scenarios, we show that our network yields improved performance and convergence rates than the baselines as the scale increases. Our results convey that IC3Net agents learn when to communicate based on the scenario and profitability.

Detecting Memorization in ReLU Networks

tl;dr We use the non-negative rank of ReLU activation matrices as a complexity measure and show it (negatively) correlates with good generalization.

We propose a new notion of 'non-linearity' of a network layer with respect to an input batch that is based on its proximity to a linear system, which is reflected in the non-negative rank of the activation matrix. We measure this non-linearity by applying non-negative factorization to the activation matrix. Considering batches of similar samples, we find that high non-linearity in deep layers is indicative of memorization. Furthermore, by applying our approach layer-by-layer, we find that the mechanism for memorization consists of distinct phases. We perform experiments on fully-connected and convolutional neural networks trained on several image and audio datasets. Our results demonstrate that as an indicator for memorization, our technique can be used to perform early stopping.

Structured Neural Summarization

tl;dr One simple trick to improve sequence models: Compose them with a graph model

Summarization of long sequences into a concise statement is a core problem in natural language processing, requiring non-trivial understanding of the input. Based on the promising results of graph neural networks on highly structured data, we develop a framework to extend existing sequence encoders with a graph component that can reason about long-distance relationships in weakly structured data such as text. In an extensive evaluation, we show that the resulting hybrid sequence-graph models outperform both pure sequence models as well as pure graph models on a range of summarization tasks.