Reproduce “Robust Classification under Noisy Labels: A Geometry-Aware Reliability Framework for Foundation Models” 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

Ecem Bozkurt, Antonio Ortega

A geometry-aware reliability framework is introduced for robust classification using pre-trained Foundation Models (FMs) with noisy labels, avoiding the need for FM retraining. This framework estimates sample trustworthiness via local and global embedding space geometry, subsequently using reliability-weighted inference to achieve improved accuracy on CIFAR-10 and DermaMNIST across various symmetric and asymmetric noise levels.

Key insights

  • Different geometric reliability strategies are optimal depending on the noise level: local, distance-based methods perform best at low noise, while global, clustering-based methods are more robust under high noise conditions.
  • The effectiveness of inference modes shifts with noise: weighted majority voting (incorporating NNK edge weights) is effective at lower noise, but unweighted voting (relying solely on reliability scores) performs better at high noise where local distances become less reliable.
  • Robustness against label noise can be significantly improved by leveraging the fixed embedding space geometry of Foundation Models without requiring their expensive retraining.

Summary by alphaXiv.

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