Reproduce “Learning to Clean: Reinforcement Learning for Noisy Label Correction” 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

Researchers from Carleton University and Amii introduce Reinforcement Learning for Noisy Label Correction (RLNLC), an approach that frames label cleaning as an active, sequential decision-making process to iteratively refine noisy datasets. RLNLC achieves state-of-the-art accuracy, showing up to 2.2% improvement on instance-dependent noise datasets like CIFAR100-IDN and a 13.2 percentage point increase on CIFAR100 with 90% symmetric noise over existing methods.

Key insights

  • Reinforcement learning's ability to make sequential, non-myopic decisions that optimize for long-term cumulative rewards is well-suited for the iterative and consequential nature of noisy label correction, overcoming the limitations of greedy, short-term approaches.
  • A carefully designed composite reward function, combining local label consistency and alignment with clean samples, effectively guides the RL policy towards discovering optimal label cleaning strategies.
  • Decoupling the policy network's feature extractor from a fixed backbone model for reward evaluation is a crucial design choice that ensures stability during RL training by isolating policy impact from label quality assessment.

Summary by alphaXiv.

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