Uncertainty-Aware Self-Paced Learning Improves Model Performance

Researchers have developed a new method called Uncertainty-Aware Self-Paced Learning (UASPL) to enhance the effectiveness of self-paced learning (SPL). Traditional SPL methods, which mimic human learning by progressing from easy to difficult samples, often rely solely on loss function values to determine sample difficulty. However, a low loss doesn't always guarantee a reliable prediction, meaning such samples aren't necessarily "easy." UASPL addresses this by integrating predictive reliability into the sample selection process. It uses evidential neural networks and a general loss function within the Subjective Logic framework, which inherently incorporates uncertainty estimation. This approach allows for a more nuanced understanding of sample difficulty and ensures interpretability in sample selection. Experimental results across various datasets demonstrate that UASPL surpasses other SPL methods in classification performance, interpretability, and overall generality.