In this paper, we show that feedforward and recurrent neural networks exhibit an outer product derivative structure but that convolutional neural networks do not. This structure makes it possible to use higher-order information without needing approximations or infeasibly large amounts of memory, and it may also provide insights into the geometry of neural network optima. The ability to easily access these derivatives also suggests a new, geometric approach to regularization. We then discuss how this structure could be used to improve training methods, increase network robustness and generalizability, and inform network compression methods.

Recently mean field theory has been successfully used to analyze properties of wide, random neural networks. It has given rise to a prescriptive theory for initializing neural networks, which ensures that the $\ell_2$ norm of the backpropagated gradients is bounded, and training is orders of magnitude faster. Despite the strong empirical performance of this class of initializations, the mechanisms by which they confer an advantage in the optimization of deep neural networks are poorly understood. Here we show a novel connection between the maximum curvature of the optimization landscape (gradient smoothness) as measured by the Fisher information matrix and the maximum singular value of the input-output Jacobian. Our theory partially explains why neural networks that are more isometric can train much faster. Furthermore, we experimentally investigate the benefits of maintaining orthogonality throughout training, from which we conclude that manifold constrained optimization of weights performs better regardless of the smoothness of the gradients. Finally we show that critical orthogonal initializations do not trivially give rise to a mean field limit of pre-activations for each layer.

We have developed an automatic sleep stage classification algorithm based on deep residual neural networks and raw polysomnogram signals. Briefly, the raw data is passed through 50 convolutional layers before subsequent classification into one of five sleep stages. Three model configurations were trained on 1850 polysomnogram recordings and subsequently tested on 230 independent recordings. Our best performing model yielded an accuracy of 84.1% and a Cohen's kappa of 0.746, improving on previous reported results by other groups also using only raw polysomnogram data. Most errors were made on non-REM stage 1 and 3 decisions, errors likely resulting from the definition of these stages. Further testing on independent cohorts is needed to verify performance for clinical use.

Historically, games of all kinds have often been the subject of study in scientific works of Computer Science, including the field of machine learning. By using machine learning techniques and applying them to a game with defined rules or a structured dataset, it's possible to learn and improve on the already existing techniques and methods to tackle new challenges and solve problems that are out of the ordinary. The already existing work on card games tends to focus on gameplay and card mechanics. This work aims to apply neural networks models, including Convolutional Neural Networks and Recurrent Neural Networks, in order to analyze Magic: the Gathering cards, both in terms of card text and illustrations; the card images and texts are used to train the networks in order to be able to classify them into multiple categories. The ultimate goal was to develop a methodology that could generate card text matching it to an input image, which was attained by relating the prediction values of the images and generated text across the different categories.

In recent years, deep neural networks have demonstrated outstanding performance in many machine learning tasks. However, researchers have discovered that these state-of-the-art models are vulnerable to adversarial examples: legitimate examples added by small perturbations which are unnoticeable to human eyes. Adversarial training, which augments the training data with adversarial examples during the training process, is a well known defense to improve the robustness of the model against adversarial attacks. However, this robustness is only effective to the same attack method used for adversarial training. Madry et al.(2017) suggest that effectiveness of iterative multi-step adversarial attacks and particularly that projected gradient descent (PGD) may be considered the universal first order adversary and applying the adversarial training with PGD implies resistance against many other first order attacks. However, the computational cost of the adversarial training with PGD and other multi-step adversarial examples is much higher than that of the adversarial training with other simpler attack techniques. In this paper, we show how strong adversarial examples can be generated only at a cost similar to that of two runs of the fast gradient sign method (FGSM), allowing defense against adversarial attacks with a robustness level comparable to that of the adversarial training with multi-step adversarial examples. We empirically demonstrate the effectiveness of the proposed two-step defense approach against different attack methods and its improvements over existing defense strategies.

The traditional approach of hand-crafting priors (such as sparsity) for solving inverse problems is slowly being replaced by the use of richer learned priors (such as those modeled by generative adversarial networks, or GANs). In this work, we study the algorithmic aspects of such a learning-based approach from a theoretical perspective. For certain generative network architectures, we establish a simple non-convex algorithmic approach that (a) theoretically enjoys linear convergence guarantees for certain inverse problems, and (b) empirically improves upon conventional techniques such as back-propagation. We also propose an extension of our approach that can handle model mismatch (i.e., situations where the generative network prior is not exactly applicable.) Together, our contributions serve as building blocks towards a more complete algorithmic understanding of generative models in inverse problems.

We present POLO --- a C++ library for large-scale parallel optimization research that emphasizes ease-of-use, flexibility and efficiency in algorithm design. It uses multiple inheritance and template programming to decompose algorithms into essential policies and facilitate code reuse. With its clear separation between algorithm and execution policies, it provides researchers with a simple and powerful platform for prototyping ideas, evaluating them on different parallel computing architectures and hardware platforms, and generating compact and efficient production code. A C-API is included for customization and data loading in high-level languages. POLO enables users to move seamlessly from serial to multi-threaded shared-memory and multi-node distributed-memory executors. We demonstrate how POLO allows users to implement state-of-the-art asynchronous parallel optimization algorithms in just a few lines of code and report experiment results from shared and distributed-memory computing architectures. We provide both POLO and POLO.jl, a wrapper around POLO written in the Julia language, at https://github.com/pologrp under the permissive MIT license.

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 nonnegative rank of 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. A fundamental challenge in machine learning is balancing the bias-variance tradeoff, where overly simple learning models underfit the data (suboptimal performance on the training data) and overly complex models are expected to overfit or memorize the data (perfect training set performance, but suboptimal test set performance). The latter direction of this tradeoff has come into question with the observation that deep neural networks do not memorize their training data despite having sufficient capacity to do so (Zhang et al., 2016), the explanation of which is a matter of much interest. Due to their convenient gradient properties and excellent performance in practice, rectified-linear units (ReLU) have been widely adopted and are now ubiquitous in the field of deep learning.

Explaining the output of a complicated machine learning model like a deep neural network (DNN) is a central challenge in machine learning. Several proposed local explanation methods address this issue by identifying what dimensions of a single input are most responsible for a DNN's output. The goal of this work is to assess the sensitivity of local explanations to DNN parameter values. Somewhat surprisingly, we find that DNNs with randomly-initialized weights produce explanations that are both visually and quantitatively similar to those produced by DNNs with learned weights. Our conjecture is that this phenomenon occurs because these explanations are dominated by the lower level features of a DNN, and that a DNN's architecture provides a strong prior which significantly affects the representations learned at these lower layers. NOTE: This work is now subsumed by our recent manuscript, Sanity Checks for Saliency Maps (to appear NIPS 2018), where we expand on findings and address concerns raised in Sundararajan et. al. (2018).

To make deep neural networks feasible in resource-constrained environments (such as mobile devices), it is beneficial to quantize models by using low-precision weights. One common technique for quantizing neural networks is the straight-through gradient method, which enables back-propagation through the quantization mapping. Despite its empirical success, little is understood about why the straight-through gradient method works. Building upon a novel observation that the straight-through gradient method is in fact identical to the well-known Nesterov's dual-averaging algorithm on a quantization constrained optimization problem, we propose a more principled alternative approach, called ProxQuant, that formulates quantized network training as a regularized learning problem instead and optimizes it via the prox-gradient method. ProxQuant does back-propagation on the underlying full-precision vector and applies an efficient prox-operator in between stochastic gradient steps to encourage quantizedness. For quantizing ResNets and LSTMs, ProxQuant outperforms state-of-the-art results on binary quantization and is on par with state-of-the-art on multi-bit quantization. For binary quantization, our analysis shows both theoretically and experimentally that ProxQuant is more stable than the straight-through gradient method (i.e. BinaryConnect), challenging the indispensability of the straight-through gradient method and providing a powerful alternative.