Reproduce “Optimal Transport-Induced Samples against Out-of-Distribution Overconfidence” 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

Keke Tang, Ziyong Du, Xiaofei Wang, Weilong Peng, Peican Zhu, Zhihong Tian

Researchers introduce Optimal Transport-Induced Samples (OTIS), a framework that leverages geometric singularities from semi-discrete optimal transport to generate semantically ambiguous out-of-distribution (OOD) samples. This method trains deep neural networks to produce less overconfident predictions on OOD inputs, outperforming existing techniques without requiring auxiliary OOD datasets.

Key insights

  • Geometric singularities in semi-discrete optimal transport maps precisely correspond to regions of semantic ambiguity and structural instability within the data manifold.
  • Optimal Transport-Induced Samples (OTIS), synthesized from these singular boundaries, effectively capture structural uncertainty and serve as powerful training signals for confidence regularization.
  • A principled, geometry-driven approach to OOD sample generation significantly outperforms heuristic methods, enhancing model trustworthiness without relying on external OOD datasets.

Summary by alphaXiv.

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