Robert Mansel Gower

Robert Mansel Gower, Nicolas Loizou, Xun Qian et al.

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.

Robert Mansel Gower, Jacek Gondzio

At the heart of Newton based optimization methods is a sequence of symmetric linear systems. Each consecutive system in this sequence is similar to the next, so solving them separately is a waste of computational effort. Here we describe automatic preconditioning techniques for iterative methods for solving such sequences of systems by maintaining an estimate of the inverse system matrix. We update the estimate of the inverse system matrix with quasi-Newton type formulas based on what we call an action constraint instead of the secant equation. We implement the estimated inverses as preconditioners in a Newton-CG method and prove quadratic termination. Our implementation is the first parallel quasi-Newton preconditioners, in full and limited memory variants. Tests on logistic Support Vector Machine problems reveal that our method is very efficient, converging in wall clock time before a Newton-CG method without preconditioning. Further tests on a set of classic test problems reveal that the method is robust. The action constraint makes these updates flexible enough to mesh with trust-region and active set methods, a flexibility that is not present in classic quasi-Newton methods.