Diverse Feature Learning by Self-distillation and Reset
This addresses feature diversity issues in machine learning models, particularly for image classification, but appears incremental as it builds on existing techniques like self-distillation and reset.
The paper tackles the problem of models struggling to learn diverse features due to forgetting or failure to learn new ones, introducing Diverse Feature Learning (DFL) that combines self-distillation for feature preservation and reset for new feature learning, resulting in identified synergistic effects in image classification experiments.
Our paper addresses the problem of models struggling to learn diverse features, due to either forgetting previously learned features or failing to learn new ones. To overcome this problem, we introduce Diverse Feature Learning (DFL), a method that combines an important feature preservation algorithm with a new feature learning algorithm. Specifically, for preserving important features, we utilize self-distillation in ensemble models by selecting the meaningful model weights observed during training. For learning new features, we employ reset that involves periodically re-initializing part of the model. As a result, through experiments with various models on the image classification, we have identified the potential for synergistic effects between self-distillation and reset.