Fuzzy Knowledge Distillation Model for Incomplete Multi-modality Data and Its Application to Brain Network Analysis

Knowledge distillation (KD) techniques have been employed to tackle the issue of incomplete multi-modality data, aiming to enhance the performance of the student model (i.e., using unimodal data) as close as possible to the teacher model (i.e., using multimodal data) through information transfer between the two models. However, directly using the traditional KD model will lead to the problem of mismatched model capacities. The teacher model typically receives a greater amount of information compared to the student model. If the learned information from the teacher model is directly transferred, it will impact the performance of the student model. To this end, this paper proposes a novel fuzzy knowledge distillation (fuzzy-KD) model for brain network analysis of incomplete multi-modality data. Specifically, we design a fuzzy-KD module to reduce the information gap between the teacher and student models through two operations: fuzzy processing and feature selection. The fuzzy processing applies an affiliation function to each channel of the feature maps learned by the teacher model to improve the generalization of representations. The feature selection employs channel compression to compel the student model to focus on crucial information within the teacher model. Experiments on real epilepsy datasets show that our method outperforms several state-of-the-art methods in the field of epilepsy recognition.