Reproduce “Depth Without the Magic: Inductive Bias of Natural Gradient Descent” 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
Anna Kerekes, Anna Mészáros, Ferenc Huszár
The paper tests whether natural gradient descent (NGD) removes training biases, comparing it with ordinary gradient descent (EGD) in sparse classification. NGD adjusts gradients using Fisher information; its invariance holds for continuous flow under qualified parameterizations. In a 1,000-feature task with 20 signal features, two- and three-layer diagonal EGD reach up to 90% test accuracy, while sample-Fisher NGD is at chance below the feature count, a setting-specific result.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.