Reproduce “AC/DC on a Budget - Alternating Sparse Phases” 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

Rahul Nittala, Advait Gadhikar, Tom Jacobs, Rebekka Burkholz

ARC performs sparse rewiring after a dense warmup, retaining active connections and adding omitted weights chosen by gradient magnitude. It trains the expanded support, then prunes to target sparsity. At 95% sparsity on ResNet50/ImageNet, ARC–Zero reports 73.087 ± 0.046% accuracy at 0.2431× dense-baseline training FLOPs, versus published AC/DC’s 73.14 ± 0.20% and 0.53×. This is not a same-run comparison, and FLOPs do not establish wall-clock acceleration.

Summary by alphaXiv.

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