Reproduce “Part-dependent Label Noise: Towards Instance-dependent Label Noise” 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

Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao and 1 more

This research introduces Part-Dependent Label Noise (PDN), a new model for handling instance-dependent label noise by approximating the instance-dependent transition matrix using a combination of part-dependent transition matrices. The proposed Part-Dependent Transition (PTD) method consistently achieves state-of-the-art classification performance on various datasets, particularly under high levels of noise, including a 10% accuracy improvement over baselines on CIFAR-10 with 50% noise and strong performance on the real-world Clothing1M dataset.

Key insights

  • Human cognitive processes, which perceive objects by first decomposing them into parts, provide a practical and intuitive basis for modeling label noise as part-dependent.
  • Instance-dependent noise, typically non-identifiable, can be effectively approximated and learned by assuming it arises from a combination of identifiable part-dependent noise characteristics.
  • Connecting the structural decomposition of an instance (via NMF) to its label noise generation process allows for a structured and more identifiable approach to IDN.

Summary by alphaXiv.

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