Self-distillation (SD) is the process of training a student model using the outputs of a teacher model, with both models sharing the same architecture. Our study theoretically examines SD in multi-class classification with cross-entropy loss, exploring both multi-round SD and SD with refined teacher outputs, inspired by partial label learning (PLL). By deriving a closed-form solution for the student model's outputs, we discover that SD essentially functions as label averaging among instances with high feature correlations. Initially beneficial, this averaging helps the model focus on feature clusters correlated with a given instance for predicting the label. However, it leads to diminishing performance with increasing distillation rounds. Additionally, we demonstrate SD's effectiveness in label noise scenarios and identify the label corruption condition and minimum number of distillation rounds needed to achieve 100% classification accuracy. Our study also reveals that one-step distillation with refined teacher outputs surpasses the efficacy of multi-step SD using the teacher's direct output in high noise rate regimes.
翻译:暂无翻译