Reproduce “Tractable structured natural gradient descent using local parameterizations” 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
Wu Lin, Frank Nielsen, Mohammad Emtiyaz Khan, Mark Schmidt
A method for tractable Natural-Gradient Descent (NGD) is presented, utilizing local parameter coordinates to address the challenges of structured parameter spaces and singular Fisher Information Matrices (FIMs). This approach systematically derives new structured second-order optimization algorithms for diverse distributions and matrix structures, achieving improved performance across various machine learning tasks.
Key insights
- The proposed framework generalizes existing NGD methods like Exponential Natural Evolution Strategies (ENES) and formally connects derived updates to Newton-like algorithms by recovering second-order terms.
- By leveraging matrix Lie groups for auxiliary parameters, the method enables the systematic derivation of novel structured second-order algorithms for various covariance structures (e.g., low-rank, diagonal, sparse, block-triangular).
- The methodology is broadly applicable beyond Gaussian distributions, successfully deriving NGD updates for Wishart distributions and showing connections to Riemannian Gradient Descent.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.