Biased Transition Matrices Break Many-Class Bottleneck in Complementary-Label Learning
The recently introduced framework, Bias-Induced Constrained Labeling (BICL), addresses the persistent scalability challenges in complementary-label learning (CLL) for multi-class scenarios. Conventional CLL approaches rely on uniform label generation, which weakens learning signals in extensive label spaces. BICL enhances learning by intentionally creating a biased (non-uniform) generation method that limits complementary labels to a specific class subset. This innovation leads to significant advancements in performance on CIFAR-100 and TinyImageNet-200, resulting in over a sevenfold increase in accuracy. The findings are available on arXiv.
Key facts
- Complementary-label learning (CLL) is a weakly supervised paradigm where instances are labeled with classes they do not belong to.
- CLL methods have been competitive mainly on 10-class classification.
- Scaling to large label spaces has been an enduring bottleneck.
- The limitation stems from the common assumption of uniform label generation.
- Uniform label generation dilutes the learning signal in many-class settings.
- BICL uses a biased (non-uniform) generation process that restricts complementary labels to a subset of classes.
- BICL achieves more than sevenfold accuracy improvement on CIFAR-100 and TinyImageNet-200.
- The paper is available on arXiv with ID 2605.15586.
Entities
Institutions
- arXiv