Simplifying Momentum-based Positive-definite Submanifold Optimization with Applications to Deep Learning
Abstract
Riemannian submanifold optimization with momentum is computationally challenging because, to ensure that the iterates remain on the submanifold, we often need to solve difficult differential equations. Here, we simplify such difficulties for a class of sparse or structured symmetric positive-definite matrices with the affine-invariant metric. We do so by proposing a generalized version of the Riemannian normal coordinates that dynamically orthonormalizes the metric and locally converts the problem into an unconstrained problem in the Euclidean space. We use our approach to simplify existing approaches for structured covariances and develop matrix-inverse-free -order optimizers for deep learning with low precision by using only matrix multiplications. Code: https://github.com/yorkerlin/StructuredNGD-DL
Cite
@article{arxiv.2302.09738,
title = {Simplifying Momentum-based Positive-definite Submanifold Optimization with Applications to Deep Learning},
author = {Wu Lin and Valentin Duruisseaux and Melvin Leok and Frank Nielsen and Mohammad Emtiyaz Khan and Mark Schmidt},
journal= {arXiv preprint arXiv:2302.09738},
year = {2024}
}
Comments
A long version of the ICML 2023 paper. Updated the main text to emphasize challenges of using existing Riemannian methods to estimate sparse and structured SPD matrices