Reproduce “Improving Context-Shift Robustness of Convolutional Networks via Context-Regularized Cross-Entropy” 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

Jinen Setpal, Chaoyue Liu

The authors introduce CR-CE, a training objective for convolutional image classifiers that reduces reliance on background context. It uses object and context masks during training and ContrastiveCAM maps to retain class evidence on objects while penalizing attribution to context. On Waterbirds, shifted-test accuracy rises from 38.40% to 79.13%, while no-shift accuracy falls from 98.48% to 94.36%; results concern evaluated benchmarks.

Summary by alphaXiv.

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