January 24: Yan Shuo Tan, Simons Institute, Berkeley

Title: Implicit regularization for constant step-size stochastic
gradient descent: why it works for non-smooth, non-convex, low-rank
models

Abstract:

Despite the development of many new optimization techniques, constant
step-size stochastic gradient descent (SGD) remains the tool of choice
in modern machine learning. It performs unreasonably well in practice,
both in convex settings where there are poor guarantees, and also in
non-convex settings, where there are often no guarantees. In this
talk, we will give an argument as to why it works for low-rank models,
illustrating the argument for the particular case of phase retrieval.

Phase retrieval is the problem of solving systems of rank-1 quadratic
equations, and can be solved by optimizing a non-convex, non-smooth
objective. In order to obtain guarantees, however, a spectral
initialization method is first used in order to get a constant error
estimate. We will show how SGD works even from arbitrary
initialization. The key insight is that the low-rank nature of the
signal allows us to study the dynamics as a Markov process with a
two-dimensional phase space, and in this phase space, the effective
step size decreases as the dimension increases. Our analysis also
introduces and makes use of several new ideas from stochastic process
theory, which we believe will be of use much more broadly.

Implicit regularization for constant step-size stochastic gradient descent: why it works for non-smooth, non-convex, low-rank models