Reproduce “Overcoming Rank Collapse in Feedback Alignment” 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
Gauthier Boeshertz, Razvan Pascanu, Claudia Clopath
The paper examines why feedback alignment, which replaces backpropagation’s shared backward weights with fixed random feedback, struggles in deeper image classifiers. It measures lower-dimensional gradients and tests Muon’s orthogonalized updates and batch normalization. On CIFAR-100 with ResNet-18, FA accuracy rises from 1.4 ± 0.0% to 46.1 ± 0.9%, remaining below BP’s 71.9 ± 0.2%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.