Reproduce “Variance Restoration over Covariance Rank: Memory-Budgeted Analytic Continual Learning” 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

Elena Voss, Pixel Nomand, Marcus Hale, Sofia Reyes

CORE-ACL compresses pairwise feature interactions while retaining each feature’s variance and full-width class-related statistics. It restores the variances before solving a full-width readout. Across five ViT-B/16 benchmarks, stage-average accuracy stays within 0.09 percentage points of dense-covariance results; on CIFAR-100, persistent state falls from 385.3 to 35.9 MiB. At equal memory, wider representations do not always win.

Summary by alphaXiv.

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