Reproduce “Efficient Adversarial Contrastive Learning via Robustness-Aware Coreset Selection” 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

Xilie Xu, Jingfeng Zhang, Feng Liu, Masashi Sugiyama, Mohan Kankanhalli

The authors introduce robustness-aware coreset selection (RCS), which selects images for adversarial contrastive learning (ACL) using gradients aligned with reductions in representational divergence (RD), an unsupervised proxy for representation robustness. On ImageNet-1K, training on a 5% subset took 111.8 hours and achieved 52.21% robust accuracy after CIFAR-10 adversarial finetuning, versus 50.12% for random selection. Full-set transfer accuracy was not reported.

Summary by alphaXiv.

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