Stochastic Mirror Descent on Overparameterized Nonlinear Models: Convergence, Implicit Regularization, and Generalization

TL;DR

This paper investigates stochastic mirror descent in overparameterized nonlinear models, showing convergence to solutions close to the initial point and demonstrating how different implicit regularizations affect generalization in deep learning.

Contribution

It provides theoretical analysis of SMD convergence properties and experimental evidence of how various implicit regularizations influence generalization performance.

Findings

SMD converges to a global minimum close to the initial point in overparameterized models.

Different SMD variants exhibit distinct generalization performances.

L10-regularized SMD outperforms SGD and L1-regularized SMD in generalization.

Abstract

Most modern learning problems are highly overparameterized, meaning that there are many more parameters than the number of training data points, and as a result, the training loss may have infinitely many global minima (parameter vectors that perfectly interpolate the training data). Therefore, it is important to understand which interpolating solutions we converge to, how they depend on the initialization point and the learning algorithm, and whether they lead to different generalization performances. In this paper, we study these questions for the family of stochastic mirror descent (SMD) algorithms, of which the popular stochastic gradient descent (SGD) is a special case. Our contributions are both theoretical and experimental. On the theory side, we show that in the overparameterized nonlinear setting, if the initialization is close enough to the manifold of global minima (something that comes for free in the highly overparameterized case), SMD with sufficiently small step size converges to a global minimum that is approximately the closest one in Bregman divergence. On the experimental side, our extensive experiments on standard datasets and models, using various initializations, various mirror descents, and various Bregman divergences, consistently confirms that this phenomenon happens in deep learning. Our experiments further indicate that there is a clear difference in the generalization performance of the solutions obtained by different SMD algorithms. Experimenting on a standard image dataset and network architecture with SMD with different kinds of implicit regularization, $\ell_1$ to encourage sparsity, $\ell_2$ yielding SGD, and $\ell_{10}$ to discourage large components in the parameter vector, consistently and definitively shows that $\ell_{10}$-SMD has better generalization performance than SGD, which in turn has better generalization performance than $\ell_1$-SMD.