Deep Linear Networks for Matrix Completion—an Infinite Depth Limit
SIAM Journal on Applied Dynamical Systems · 2023-12-31 · 22 (4) · 3208-3232
DOI: 10.1137/22m1530653 ↗At a glance
Journal article in SIAM Journal on Applied Dynamical Systems, published online on 2023-11-28. Written by Nadav Cohen and 2 co-authors. Cited 6 times so far (Crossref); 21 references. Funded by: Israel Science Foundation, Division of Mathematical Sciences.
Abstract (original language)
The deep linear network (DLN) is a model for implicit regularization in gradient based optimization of overparametrized learning architectures. Training the DLN corresponds to a Riemannian gradient flow, where the Riemannian metric is defined by the architecture of the network and the loss function is defined by the learning task. We extend this geometric framework, obtaining explicit expressions for the volume form, including the case when the network has infinite depth. We investigate the link between the Riemannian geometry and the training asymptotics for matrix completion with rigorous analysis and numerics. We propose that under small initialization, implicit regularization is a result of bias towards high state space volume.
Funding
- Israel Science Foundation
- Division of Mathematical Sciences
Cite
APA
Cohen, N., Menon, G., & Veraszto, Z. (2023). Deep Linear Networks for Matrix Completion—an Infinite Depth Limit. SIAM Journal on Applied Dynamical Systems, 22(4), 3208-3232. https://doi.org/10.1137/22m1530653
Vancouver
Cohen N, Menon G, Veraszto Z. Deep Linear Networks for Matrix Completion—an Infinite Depth Limit. SIAM Journal on Applied Dynamical Systems. 2023;22(4):3208-3232. doi:10.1137/22m1530653
BibTeX
@article{Cohen2023,
author = {Cohen, Nadav and Menon, Govind and Veraszto, Zsolt},
title = {Deep Linear Networks for Matrix Completion—an Infinite Depth Limit},
journal = {SIAM Journal on Applied Dynamical Systems},
year = {2023},
volume = {22},
number = {4},
pages = {3208--3232},
doi = {10.1137/22m1530653}
}
EUVIMED