Reproduce “DivideMix: Learning with Noisy Labels as Semi-supervised Learning” 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
Junnan Li, Richard Socher, Steven C. H. Hoi
Salesforce Research developed DivideMix, a framework that addresses learning with noisy labels by re-conceptualizing noisy samples as unlabeled data within a semi-supervised learning paradigm. The method achieves state-of-the-art accuracy, showing improvements of up to 12% absolute on real-world datasets like WebVision and maintaining robust performance even at 90% label noise.
Key insights
- Noisy labeled samples, despite their incorrect labels, still contain valuable feature information that can be effectively utilized as "unlabeled data" within a semi-supervised learning context.
- Employing a co-training mechanism with two independent networks that mutually inform data partitioning and pseudo-labeling is critical for preventing confirmation bias and ensuring robust learning from noisy data.
- Combining dynamic data division with consistency regularization, label refinement, and ensemble-based pseudo-labeling (inspired by MixMatch) creates a powerful synergy for handling various types and levels of label noise effectively.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.