Weakly Supervised Learning Advances for Data Scarcity.

Deep learning's success heavily relies on abundant, high-quality annotated training data, a requirement often unmet in real-world scenarios. Weakly supervised learning (WSL) aims to overcome this by enabling the training of accurate models using less-than-ideal supervision, which can be incomplete, inexact, or inaccurate. This review highlights recent progress in the field of WSL. It introduces novel supervision paradigms, such as "confidence-difference classification," and proposes consistent methods to tackle this binary classification problem. The discussion also covers "complementary-label learning," a multi-class classification challenge, presenting new approaches that operate under more relaxed data generation assumptions than previous consistent methods. Furthermore, the chapter addresses "partial-label learning," another prevalent multi-class WSL problem. To foster fair and realistic algorithm evaluation in this area, it proposes a new evaluation framework. These advancements collectively offer practical solutions for developing robust AI models even when high-quality labeled data is scarce.