Reproduce “Reduction from Complementary-Label Learning to Probability Estimates” 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
Wei-I Lin, Hsuan-Tien Lin
A novel framework for Complementary-Label Learning (CLL) is proposed, which reduces the problem to estimating complementary class probabilities instead of directly reducing to ordinary classification. This paradigm enables greater flexibility in model choice, exhibits robustness against noisy and biased complementary labels, and achieves strong empirical performance across various datasets and conditions.
Key insights
- Minimizing the KL divergence between estimated and observed complementary label distributions is theoretically equivalent to minimizing the divergence from the true complementary label distribution (P(ȳ|x,y)), enabling training with only complementary labels.
- Theoretical bounds demonstrate that accurate complementary probability estimates, combined with L1 decoding, directly lead to accurate ordinary classification and that classification error gracefully degrades with inaccuracies in the transition matrix.
- Several existing state-of-the-art CLL methods (e.g., Forward Correction, Surrogate Complementary Loss with log-loss, Discriminative Models) can be re-interpreted as special cases within the proposed CPE framework, providing a unifying theoretical perspective.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.