Reproduce “Reinforcement Learning-Guided 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

Marzi Heidari, Hanping Zhang, Yuhong Guo

The paper introduces Reinforcement Learning-Guided Semi-Supervised Learning (RLGSSL), a method that formulates SSL as a one-armed bandit problem to dynamically guide the learning process. RLGSSL achieves lower test error rates compared to state-of-the-art SSL techniques across image classification benchmarks like CIFAR-10, CIFAR-100, and SVHN, under various labeled data settings and network architectures.

Key insights

  • Formulating SSL as a one-armed bandit problem enables the prediction model to dynamically adapt its learning strategy, moving beyond heuristic-driven approaches for utilizing unlabeled data.
  • A reward function based on data mixup and negative Mean Squared Error, combined with a KL-divergence weighted RL loss, encourages discriminative, accurate, and robust predictions.
  • The synergistic integration of RL guidance with a stable EMA-based teacher model and foundational SSL losses creates a comprehensive and highly effective learning objective for semi-supervised tasks.

Summary by alphaXiv.

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