Characterizing Implicit Bias in Terms of Optimization Geometry
Abstract
We study the implicit bias of generic optimization methods, such as mirror descent, natural gradient descent, and steepest descent with respect to different potentials and norms, when optimizing underdetermined linear regression or separable linear classification problems. We explore the question of whether the specific global minimum (among the many possible global minima) reached by an algorithm can be characterized in terms of the potential or norm of the optimization geometry, and independently of hyperparameter choices such as step-size and momentum.
Cite
@article{arxiv.1802.08246,
title = {Characterizing Implicit Bias in Terms of Optimization Geometry},
author = {Suriya Gunasekar and Jason Lee and Daniel Soudry and Nathan Srebro},
journal= {arXiv preprint arXiv:1802.08246},
year = {2020}
}
Comments
(1) A bug in the proof of implicit bias for matrix factorization was fixed. v2 gives a characterization of the asymptotic bias of the factor matrices, while v1 made a stronger claim on the limit direction of the unfactored matrix. (2) v2 also includes new results on implicit bias of mirror descent with realizable affine constraints