Reproduce “Top-KAST: Top-K Always Sparse Training” 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
Siddhant M. Jayakumar, Razvan Pascanu, Jack W. Rae, Simon Osindero, Erich Elsen
Top-KAST trains sparse networks by using the largest-magnitude weights for predictions while updating a larger set of candidate weights, allowing connections to change. On ImageNet ResNet-50, magnitude selection improves accuracy over random candidates at 95% forward and 90% backward sparsity, but not at 90% and 80%. Its language results are competitive, but do not establish wall-clock or energy savings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.