It is well known that many optimization methods, including SGD, SAGA, and Accelerated SGD for over-parameterized models, do not scale linearly in the parallel setting. In this paper, we present a new version of block coordinate descent that solves this issue for a number of methods. The core idea is to make the sampling of coordinate blocks on each parallel unit independent of the others. Surprisingly, we prove that the optimal number of blocks to be updated by each of $n$ units in every iteration is equal to $m/n$, where $m$ is the total number of blocks. As an illustration, this means that when $n=100$ parallel units are used, $99\%$ of work is a waste of time. We demonstrate that with $m/n$ blocks used by each unit the iteration complexity often remains the same. Among other applications which we mention, this fact can be exploited in the setting of distributed optimization to break the communication bottleneck. Our claims are justified by numerical experiments which demonstrate almost a perfect match with our theory on a number of datasets.

Learning representations that disentangle the underlying factors of variability in data is an intuitive precursor to AI with human-like reasoning. Consequently, it has been the object of many efforts of the machine learning community. This work takes a step further in this direction by addressing the scenario where generative factors present a multimodal distribution due to the existence of class distinction in the data. We formulate a lower bound on the joint distribution of inputs and class labels and present N-VAE, a model which is capable of separating factors of variation which are exclusive to certain classes from factors that are shared among classes. This model implements the natural clustering prior through the use of a class-conditioned latent space and a shared latent space. We show its usefulness for detecting and disentangling class-dependent generative factors as well as for generating rich artificial samples.

We propose a general yet simple theorem describing the convergence of SGD under the arbitrary sampling paradigm. Our theorem describes the convergence of an infinite array of variants of SGD, each of which is associated with a specific probability law governing the data selection rule used to form mini-batches. This is the first time such an analysis is performed, and most of our variants of SGD were never explicitly considered in the literature before. Our analysis relies on the recently introduced notion of expected smoothness and does not rely on a uniform bound on the variance of the stochastic gradients. By specializing our theorem to different mini-batching strategies, such as sampling with replacement and independent sampling, we derive exact expressions for the stepsize as a function of the mini-batch size. With this we can also determine the mini-batch size that optimizes the total complexity, and show explicitly that as the variance of the stochastic gradient evaluated at the minimum grows, so does the optimal mini-batch size. For zero variance, the optimal mini-batch size is one. Moreover, we prove insightful stepsize-switching rules which describe when one should switch from a constant to a decreasing stepsize regime.

We propose a simple objective evaluation measure for explanations of a complex black-box machine learning model. While most such model explanations have largely been evaluated via qualitative measures, such as how humans might qualitatively perceive the explanations, it is vital to also consider objective measures such as the one we propose in this paper. Our evaluation measure that we naturally call sensitivity is simple: it characterizes how an explanation changes as we vary the test input, and depending on how we measure these changes, and how we vary the input, we arrive at different notions of sensitivity. We also provide a calculus for deriving sensitivity of complex explanations in terms of that for simpler explanations, which thus allows an easy computation of sensitivities for yet to be proposed explanations. One advantage of an objective evaluation measure is that we can optimize the explanation with respect to the measure: we show that (1) any given explanation can be simply modified to improve its sensitivity with just a modest deviation from the original explanation, and (2) gradient based explanations of an adversarially trained network are less sensitive. Perhaps surprisingly, our experiments show that explanations optimized to have lower sensitivity can be more faithful to the model predictions.

Stochastic approximation (SA) is a classical approach for stochastic convex optimization. Previous studies have demonstrated that the convergence rate of SA can be improved by introducing either smoothness or strong convexity condition. In this paper, we make use of smoothness and strong convexity simultaneously to boost the convergence rate. Let $\lambda$ be the modulus of strong convexity, $\kappa$ be the condition number, $F_*$ be the minimal risk, and $\alpha>1$ be some small constant. First, we demonstrate that, in expectation, an $O(1/[\lambda T^\alpha] + \kappa F_*/T)$ risk bound is attainable when $T = \Omega(\kappa^\alpha)$. Thus, when $F_*$ is small, the convergence rate could be faster than $O(1/[\lambda T])$ and approaches $O(1/[\lambda T^\alpha])$ in the ideal case. Second, to further benefit from small risk, we show that, in expectation, an $O(1/2^{T/\kappa}+F_*)$ risk bound is achievable. Thus, the excess risk reduces exponentially until reaching $O(F_*)$, and if $F_*=0$, we obtain a global linear convergence. Finally, we emphasize that our proof is constructive and each risk bound is equipped with an efficient stochastic algorithm attaining that bound.

Large-batch SGD is important for scaling training of deep neural networks. However, without fine-tuning hyperparameter schedules, the generalization of the model may be hampered. We propose to use batch augmentation: replicating instances of samples within the same batch with different data augmentations. Batch augmentation acts as a regularizer and an accelerator, increasing both generalization and performance scaling. We analyze the effect of batch augmentation on gradient variance and show that it empirically improves convergence for a wide variety of deep neural networks and datasets. Our results show that batch augmentation reduces the number of necessary SGD updates to achieve the same accuracy as the state-of-the-art. Overall, this simple yet effective method enables faster training and better generalization by allowing more computational resources to be used concurrently.

Reward shaping is one of the most effective methods to tackle the crucial yet challenging problem of credit assignment in Reinforcement Learning (RL). However, designing shaping functions usually requires much expert knowledge and hand-engineering, and the difficulties are further exacerbated given multiple similar tasks to solve. In this paper, we consider reward shaping on a distribution of tasks, and propose a general meta-learning framework to automatically learn the efficient reward shaping on newly sampled tasks, assuming only shared state space but not necessarily action space. We first derive the theoretically optimal reward shaping in terms of credit assignment in model-free RL. We then propose a value-based meta-learning algorithm to extract an effective prior over the optimal reward shaping. The prior can be applied directly to new tasks, or provably adapted to the task-posterior while solving the task within few gradient updates. We demonstrate the effectiveness of our shaping through significantly improved learning efficiency and interpretable visualizations across various settings, including notably a successful transfer from DQN to DDPG.

This paper takes aim at achieving nothing less than the impossible. To be more precise, we seek to predict labels of unknown data from entirely uncorrelated labelled training data. This will be accomplished by an application of an algorithm based on deep learning, as well as, by invoking one of the most fundamental concepts of set theory. Estimating the behaviour of a system in unknown situations is one of the central problems of humanity. Indeed, we are constantly trying to produce predictions for future events to be able to prepare ourselves.

In this paper we propose a Bayesian method for estimating architectural parameters of neural networks, namely layer size and network depth. We do this by learning concrete distributions over these parameters. Our results show that regular networks with a learnt structure can generalise better on small datasets, while fully stochastic networks can be more robust to parameter initialisation. The proposed method relies on standard neural variational learning and, unlike randomised architecture search, does not require a retraining of the model, thus keeping the computational overhead at minimum.

Determinantal point processes (DPPs) are probabilistic models for repulsion. When used to represent the occurrence of random subsets of a finite base set, DPPs allow to model global negative associations in a mathematically elegant and direct way. Discrete DPPs have become popular and computationally tractable models for solving several machine learning tasks that require the selection of diverse objects, and have been successfully applied in numerous real-life problems. Despite their popularity, the statistical properties of such models have not been adequately explored. In this note, we derive the Markov properties of discrete DPPs and show how they can be expressed using graphical models.