Reproduce “Tilted Empirical Risk Minimization” 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

Tian Li, Ahmad Beirami, Maziar Sanjabi, Virginia Smith

TERM replaces ERM’s average training loss with an exponential tilt, reweighting examples by loss and optionally tilting classes or groups. Zero tilt recovers ERM; positive and negative tilts change priorities. In HIV-1’s hardest tested condition, TERM_ca achieved 0.806 rare-class and 0.901 overall accuracy with 30% reshuffled labels and a 1:20 ratio, but this was random noise without a hierarchy ablation.

Summary by alphaXiv.

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