Reproduce “FlexMatch: Boosting Semi-Supervised Learning with Curriculum Pseudo Labeling” 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

Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, Takahiro Shinozaki

FlexMatch, developed by researchers from Tokyo Institute of Technology and Microsoft, introduces Curriculum Pseudo Labeling (CPL) to dynamically adjust pseudo-labeling confidence thresholds per class. This approach boosts semi-supervised learning accuracy and convergence speed, particularly in low-label settings, surpassing FixMatch on most benchmarks.

Key insights

  • Dynamically adjusting pseudo-labeling confidence thresholds for each class, based on its current learning status, is more effective than a global fixed threshold for semi-supervised learning.
  • The learning status of individual classes can be accurately and efficiently approximated by tracking the count of high-confidence pseudo-labeled samples assigned to each class.
  • Incorporating a threshold warm-up period and a convex non-linear mapping function for threshold adjustment stabilizes early training and optimizes the curriculum progression, allowing thresholds to adapt more effectively.

Summary by alphaXiv.

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