Embracing Biased Transition Matrices for Complementary-Label Learning with Many Classes

In three linesComplementary-label learning (CLL) methods remain limited to 10-class classification. This paper proposes BICL, a framework that deliberately uses biased (non-uniform) transition matrices to restrict complementary labels to class subsets. On CIFAR-100 and TinyImageNet-200, BICL achieves 7× accuracy improvements over traditional methods.

Summary generated by Claude — human-verified