Knowledge Distillation via Information Matching

Honglin Zhu

Abstract: Knowledge distillation can enhance network generalization by guiding a smaller student network to learn from a more complex teacher network. The challenge lies in maximizing the performance of the student network under the supervision of the teacher network. Currently, the feature-based distillation approach utilizes the middle-layer features of the teacher network to improve the performance of the student network. However, this approach lacks a measure to evaluate the content of the information present in the intermediate layers of both the teacher and student networks, which leads to a distillation mismatch of features and damages the student’s performance. In this study, we propose a new feature distillation method to solve this problem. We measure the information content in the intermediate layers of the teacher and student networks based on the receptive fields of corresponding features. Subsequently, the suitable number and locations of transmission features are decided based on information content, effectively alleviating the risk of information mismatch during distillation. Our experimental results demonstrate that the proposed method significantly improves the performance of the student network.