Reproduce “Weakly-Supervised Contrastive Learning for Imprecise Class Labels” 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
Zi-Hao Zhou, Jun-Jie Wang, Tong Wei, Min-Ling Zhang
This research from Southeast University introduces a Weakly-Supervised Contrastive (WSC) learning framework that leverages "continuous semantic similarity" and a graph-theoretic approach to enable supervised contrastive learning even with imprecise class labels. It achieves competitive or superior classification accuracy on simulated and real-world noisy and partial label datasets, outperforming previous methods by up to 7.8% on CIFAR-100 with a 0.3 partial ratio and 6.85% on CIFAR-100 with 90% symmetric noise.
Key insights
- The framework offers a unified and versatile approach to handle various forms of weak supervision (noisy labels, partial labels) by generalizing the notion of positive/negative pairs with continuous semantic similarity.
- Under mild conditions, the proposed WSC framework theoretically approximates the performance achievable with fully supervised information, establishing an error bound for the learned features.
- The integration of self-supervised and weakly-supervised connectivity in the perturbation graph effectively enhances intra-class density, improving clustering quality even with imprecise labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.