Reproduce “Noisy-Pair Robust Representation Alignment for Positive-Unlabeled 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
Hengwei Zhao, Zhengzhong Tu, Zhuo Zheng, Wei Wang, Junjue Wang, Rusty Feagin, Wenzhe Jiao
Researchers at Texas A&M University and collaborators developed NcPU, a Positive-Unlabeled (PU) learning framework that acquires discriminative representations from limited positive and abundant unlabeled data without auxiliary information. NcPU consistently achieved the highest performance across various datasets, effectively closing the gap with fully supervised methods by addressing noisy supervision.
Key insights
- Standard supervised non-contrastive loss can be misled by noisy pairs in PU settings, where gradients from incorrect pairings can overpower those from correct ones, hindering discriminative representation learning.
- The iterative interplay between robust representation learning (NoiSNCL) and refined label disambiguation (PLD) mutually enhances performance, functioning effectively within an Expectation-Maximization (EM) framework.
- Achieving high-quality, discriminative representations is a critical factor for improving PU learning algorithms, enabling even simpler risk estimation methods to perform comparably to complex state-of-the-art approaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.