Arguably the biggest challenge in applying neural networks is tuning the hyperparameters, in particular the learning rate. The sensitivity to the learning rate is due to the reliance on backpropagation to train the network. In this paper we present the first application of Implicit Stochastic Gradient Descent (ISGD) to train neural networks, a method known in convex optimization to be unconditionally stable and robust to the learning rate. Our key contribution is a novel layer-wise approximation of ISGD which makes its updates tractable for neural networks. Experiments show that our method is more robust to high learning rates and generally outperforms standard backpropagation on a variety of tasks.

In this paper we study a family of variance reduction methods with randomized batch size---at each step, the algorithm first randomly chooses the batch size and then selects a batch of samples to conduct a variance-reduced stochastic update. We give the linear convergence rate for this framework for composite functions, and show that the optimal strategy to achieve the optimal convergence rate per data access is to always choose batch size of 1, which is equivalent to the SAGA algorithm. However, due to the presence of cache/disk IO effect in computer architecture, the number of data access cannot reflect the running time because of 1) random memory access is much slower than sequential access, 2) when data is too big to fit into memory, disk seeking takes even longer time. After taking these into account, choosing batch size of $1$ is no longer optimal, so we propose a new algorithm called SAGA++ and show how to calculate the optimal average batch size theoretically. Our algorithm outperforms SAGA and other existing batched and stochastic solvers on real datasets. In addition, we also conduct a precise analysis to compare different update rules for variance reduction methods, showing that SAGA++ converges faster than SVRG in theory.

Furthermore, traveling in any direction orthogonal to the gradient maintains the value of the function. In this work, we show that these orthogonal directions that are ignored by gradient descent can be critical in equilibrium problems. Equilibrium problems have drawn heightened attention in machine learning due to the emergence of the Generative Adversarial Network (GAN). We use the framework of Variational Inequalities to analyze popular training algorithms for a fundamental GAN variant: the Wasserstein Linear-Quadratic GAN. We show that the steepest descent direction causes divergence from the equilibrium, and guaranteed convergence to the equilibrium is achieved through following a particular orthogonal direction. We call this successful technique Crossing-the-Curl, named for its mathematical derivation as well as its intuition: identify the game's axis of rotation and move "across" space in the direction towards smaller "curling".

Neural networks have many successful applications, while much less theoretical understanding has been gained. Towards bridging this gap, we study the problem of learning a two-layer overparameterized ReLU neural network for multi-class classification via stochastic gradient descent (SGD) from random initialization. In the overparameterized setting, when the data comes from mixtures of well-separated distributions, we prove that SGD learns a network with a small generalization error, albeit the network has enough capacity to fit arbitrary labels. Furthermore, the analysis provides interesting insights into several aspects of learning neural networks and can be verified based on empirical studies on synthetic data and on the MNIST dataset.

Bayesian inference plays an important role in advancing machine learning, but faces computational challenges when applied to complex models such as deep neural networks. Variational inference circumvents these challenges by formulating Bayesian inference as an optimization problem and solving it using gradient-based optimization. In this paper, we argue in favor of natural-gradient approaches which, unlike their gradient-based counterparts, can improve convergence by exploiting the information geometry of the solutions. We show how to derive fast yet simple natural-gradient updates by using a duality associated with exponential-family distributions. An attractive feature of these methods is that, by using natural-gradients, they are able to extract accurate local approximations for individual model components. We summarize recent results for Bayesian deep learning showing the superiority of natural-gradient approaches over their gradient counterparts.

Stochastic gradient descent (SGD), which dates back to the 1950s, is one of the most popular and effective approaches for performing stochastic optimization. Research on SGD resurged recently in machine learning for optimizing convex loss functions as well as training nonconvex deep neural networks. The theory assumes that one can easily compute an unbiased gradient estimator, which is usually the case due to the sample average nature of empirical risk minimization. There exist, however, many scenarios (e.g., graph learning) where an unbiased estimator may be as expensive to compute as the full gradient, because training examples are interconnected. In a recent work, Chen et al. (2018) proposed using a consistent gradient estimator as an economic alternative. Encouraged by empirical success, we show, in a general setting, that consistent estimators result in the same convergence behavior as do unbiased ones. Our analysis covers strongly convex, convex, and nonconvex objectives. This work opens several new research directions, including the development of more efficient SGD updates with consistent estimators and the design of efficient training algorithms for large-scale graphs.

The use of complex models --with many parameters-- is challenging with high-dimensional small-sample problems: indeed, they face rapid overfitting. Such situations are common when data collection is expensive, as in neuroscience, biology, or geology. Dedicated regularization can be crafted to tame overfit, typically via structured penalties. But rich penalties require mathematical expertise and entail large computational costs. Stochastic regularizers such as dropout are easier to implement: they prevent overfitting by random perturbations. Used inside a stochastic optimizer, they come with little additional cost. We propose a structured stochastic regularization that relies on feature grouping. Using a fast clustering algorithm, we define a family of groups of features that capture feature covariations. We then randomly select these groups inside a stochastic gradient descent loop. This procedure acts as a structured regularizer for high-dimensional correlated data without additional computational cost and it has a denoising effect. We demonstrate the performance of our approach for logistic regression both on a sample-limited face image dataset with varying additive noise and on a typical high-dimensional learning problem, brain image classification.

This paper introduces Jensen, an easily extensible and scalable toolkit for production-level machine learning and convex optimization. Jensen implements a framework of convex (or loss) functions, convex optimization algorithms (including Gradient Descent, L-BFGS, Stochastic Gradient Descent, Conjugate Gradient, etc.), and a family of machine learning classifiers and regressors (Logistic Regression, SVMs, Least Square Regression, etc.). This framework makes it possible to deploy and train models with a few lines of code, and also extend and build upon this by integrating new loss functions and optimization algorithms.

By and large the behavior of stochastic gradient is regarded as a challenging problem, and it is often presented in the framework of statistical machine learning. This paper offers a novel view on the analysis of on-line models of learning that arises when dealing with a generalized version of stochastic gradient that is based on dissipative dynamics. In order to face the complex evolution of these models, a systematic treatment is proposed which is based on energy balance equations that are derived by means of the Caldirola-Kanai (CK) Hamiltonian. According to these equations, learning can be regarded as an ordering process which corresponds with the decrement of the loss function. Finally, the main results established in this paper is that in the case of quasi-periodic environments, where the pattern novelty is progressively limited as time goes by, the system dynamics yields an asymptotically consistent solution in the weight space, that is the solution maps similar patterns to the same decision.

Games generalize the optimization paradigm by introducing different objective functions for different optimizing agents, known as players. Generative Adversarial Networks (GANs) are arguably the most popular game formulation in recent machine learning literature. GANs achieve great results on generating realistic natural images, however they are known for being difficult to train. Training them involves finding a Nash equilibrium, typically performed using gradient descent on the two players' objectives. Game dynamics can induce rotations that slow down convergence to a Nash equilibrium, or prevent it altogether. We provide a theoretical analysis of the game dynamics. Our analysis, supported by experiments, shows that gradient descent with a negative momentum term can improve the convergence properties of some GANs.