Variational autoencoders (VAEs) are powerful generative models with the salient ability to perform inference. Here, we introduce a \emph{quantum variational autoencoder} (QVAE): a VAE whose latent generative process is implemented as a quantum Boltzmann machine (QBM). We show that our model can be trained end-to-end by maximizing a well-defined loss-function: a "quantum" lower-bound to a variational approximation of the log-likelihood. We use quantum Monte Carlo (QMC) simulations to train and evaluate the performance of QVAEs. To achieve the best performance, we first create a VAE platform with discrete latent space generated by a restricted Boltzmann machine (RBM). Our model achieves state-of-the-art performance on the MNIST dataset when compared against similar approaches that only involve discrete variables in the generative process. We consider QVAEs with a smaller number of latent units to be able to perform QMC simulations, which are computationally expensive. We show that QVAEs can be trained effectively in regimes where quantum effects are relevant despite training via the quantum bound. Our findings open the way to the use of quantum computers to train QVAEs to achieve competitive performance for generative models. Placing a QBM in the latent space of a VAE leverages the full potential of current and next-generation quantum computers as sampling devices.

Clinical notes are text documents that are created by clinicians for each patient encounter. They are typically accompanied by medical codes, which describe the diagnosis and treatment. Annotating these codes is labor intensive and error prone; furthermore, the connection between the codes and the text is not annotated, obscuring the reasons and details behind specific diagnoses and treatments. We present an attentional convolutional network that predicts medical codes from clinical text. Our method aggregates information across the document using a convolutional neural network, and uses an attention mechanism to select the most relevant segments for each of the thousands of possible codes. The method is accurate, achieving precision @ 8 of 0.7 and a Micro-F1 of 0.52, which are both significantly better than the prior state of the art. Furthermore, through an interpretability evaluation by a physician, we show that the attention mechanism identifies meaningful explanations for each code assignment.

This paper investigates recently proposed approaches for defending against adversarial examples and evaluating adversarial robustness. The existence of adversarial examples in trained neural networks reflects the fact that expected risk alone does not capture the model's performance against worst-case inputs. We motivate the use of adversarial risk as an objective, although it cannot easily be computed exactly. We then frame commonly used attacks and evaluation metrics as defining a tractable surrogate objective to the true adversarial risk. This suggests that models may be obscured to adversaries, by optimizing this surrogate rather than the true adversarial risk. We demonstrate that this is a significant problem in practice by repurposing gradient-free optimization techniques into adversarial attacks, which we use to decrease the accuracy of several recently proposed defenses to near zero. Our hope is that our formulations and results will help researchers to develop more powerful defenses.

Convolutional operator learning is increasingly gaining attention in many signal processing and computer vision applications. Learning kernels has mostly relied on so-called local approaches that extract and store many overlapping patches across training signals. Due to memory demands, local approaches have limitations when learning kernels from large datasets -- particularly with multi-layered structures, e.g., convolutional neural network (CNN) -- and/or applying the learned kernels to high-dimensional signal recovery problems. The so-called global approach has been studied within the "synthesis" signal model, e.g., convolutional dictionary learning, overcoming the memory problems by careful algorithmic designs. This paper proposes a new convolutional analysis operator learning (CAOL) framework in the global approach, and develops a new convergent Block Proximal Gradient method using a Majorizer (BPG-M) to solve the corresponding block multi-nonconvex problems. To learn diverse filters within the CAOL framework, this paper introduces an orthogonality constraint that enforces a tight-frame (TF) filter condition, and a regularizer that promotes diversity between filters. Numerical experiments show that, for tight majorizers, BPG-M significantly accelerates the CAOL convergence rate compared to the state-of-the-art method, BPG. Numerical experiments for sparse-view computational tomography show that CAOL using TF filters significantly improves reconstruction quality compared to a conventional edge-preserving regularizer. Finally, this paper shows that CAOL can be useful to mathematically model a CNN, and the corresponding updates obtained via BPG-M coincide with core modules of the CNN.

Recent works investigated the generalization properties in deep neural networks (DNNs) by studying the Information Bottleneck in DNNs. However, the mea- surement of the mutual information (MI) is often inaccurate due to the density estimation. To address this issue, we propose to measure the dependency instead of MI between layers in DNNs. Specifically, we propose to use Hilbert-Schmidt Independence Criterion (HSIC) as the dependency measure, which can measure the dependence of two random variables without estimating probability densities. Moreover, HSIC is a special case of the Squared-loss Mutual Information (SMI). In the experiment, we empirically evaluate the generalization property using HSIC in both the reconstruction and prediction auto-encoding (AE) architectures.

Momentum Stochastic Gradient Descent (MSGD) algorithm has been widely applied to many nonconvex optimization problems in machine learning. Popular examples include training deep neural networks, dimensionality reduction, and etc. Due to the lack of convexity and the extra momentum term, the optimization theory of MSGD is still largely unknown. In this paper, we study this fundamental optimization algorithm based on the so-called "strict saddle problem." By diffusion approximation type analysis, our study shows that the momentum helps escape from saddle points, but hurts the convergence within the neighborhood of optima (if without the step size annealing). Our theoretical discovery partially corroborates the empirical success of MSGD in training deep neural networks. Moreover, our analysis applies the martingale method and "Fixed-State-Chain" method from the stochastic approximation literature, which are of independent interest.

Deep neural networks have enabled progress in a wide variety of applications. Growing the size of the neural network typically results in improved accuracy. As model sizes grow, the memory and compute requirements for training these models also increases. We introduce a technique to train deep neural networks using half precision floating point numbers. In our technique, weights, activations and gradients are stored in IEEE half-precision format. Half-precision floating numbers have limited numerical range compared to single-precision numbers. We propose two techniques to handle this loss of information. Firstly, we recommend maintaining a single-precision copy of the weights that accumulates the gradients after each optimizer step. This single-precision copy is rounded to half-precision format during training. Secondly, we propose scaling the loss appropriately to handle the loss of information with half-precision gradients. We demonstrate that this approach works for a wide variety of models including convolution neural networks, recurrent neural networks and generative adversarial networks. This technique works for large scale models with more than 100 million parameters trained on large datasets. Using this approach, we can reduce the memory consumption of deep learning models by nearly 2x. In future processors, we can also expect a significant computation speedup using half-precision hardware units.

Face recognition is a widely used technology with numerous large-scale applications, such as surveillance, social media and law enforcement. There has been tremendous progress in face recognition accuracy over the past few decades, much of which can be attributed to deep learning-based approaches during the last five years. Indeed, automated face recognition systems are now believed to surpass human performance in some scenarios. Despite this progress, a crucial question still remains unanswered: given a face representation, how many identities can it resolve? In other words, what is the capacity of the face representation? A scientific basis for estimating the capacity of a given face representation will not only benefit the evaluation and comparison of different face representations but will also establish an upper bound on the scalability of an automatic face recognition system. We cast the face capacity estimation problem under the information theoretic framework of capacity of a Gaussian noise channel. By explicitly accounting for two sources of representational noise: epistemic uncertainty and aleatoric variability, our approach is able to estimate the capacity of any given face representation. To demonstrate the efficacy of our approach, we estimate the capacity of a 128-dimensional DNN based face representation, FaceNet, and that of the classical Eigenfaces representation of the same dimensionality. Our experiments on unconstrained faces indicate that, (a) our proposed model yields a capacity upper bound of 5.8x$10^{8}$ for FaceNet and 1x$10^{0}$ for Eigenfaces at a false acceptance rate (FAR) of 1%, (b) the face representation capacity reduces drastically as you lower the desired FAR (for FaceNet; the capacity at FAR of 0.1% and 0.001% is 2.4x$10^{6}$ and 7.0x$10^{2}$, respectively), and (c) the empirical performance of FaceNet is significantly below the theoretical limit.

We consider variants of trust-region and cubic regularization methods for non-convex optimization, in which the Hessian matrix is approximated. Under mild conditions on the inexact Hessian, and using approximate solution of the corresponding sub-problems, we provide iteration complexity to achieve $ \epsilon $-approximate second-order optimality which have shown to be tight. Our Hessian approximation conditions constitute a major relaxation over the existing ones in the literature. Consequently, we are able to show that such mild conditions allow for the construction of the approximate Hessian through various random sampling methods. In this light, we consider the canonical problem of finite-sum minimization, provide appropriate uniform and non-uniform sub-sampling strategies to construct such Hessian approximations, and obtain optimal iteration complexity for the corresponding sub-sampled trust-region and cubic regularization methods.

We introduce NoisyNet, a deep reinforcement learning agent with parametric noise added to its weights, and show that the induced stochasticity of the agent's policy can be used to aid efficient exploration. The parameters of the noise are learned with gradient descent along with the remaining network weights. NoisyNet is straightforward to implement and adds little computational overhead. We find that replacing the conventional exploration heuristics for A3C, DQN and dueling agents (entropy reward and $\epsilon$-greedy respectively) with NoisyNet yields substantially higher scores for a wide range of Atari games, in some cases advancing the agent from sub to super-human performance.