Reproduce “The Law of Parsimony in Gradient Descent for Learning Deep Linear Networks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Can Yaras, Peng Wang, Wei Hu, Zhihui Zhu, Laura Balzano, Qing Qu

Researchers at the University of Michigan demonstrated a "law of parsimony" in deep linear networks trained with gradient descent, where learning primarily occurs within minimal invariant subspaces of weight matrices. This principle explains the implicit bias towards low-rank solutions and enables the creation of computationally efficient "compressed" networks.

Key insights

  • Gradient Descent implicitly restricts learning to a minimal invariant subspace within each weight matrix of DLNs, leaving a large portion of parameters effectively dormant.
  • This parsimonious learning dynamics provides a theoretical explanation for the low-rank bias observed in deep networks, reinforced by weight decay causing unused singular values to diminish.
  • The "law of parsimony" explains how deep layers progressively refine data representations by linearly decaying a data separation metric, leading to feature concentration within classes.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.