Multi-view Teacher-Student Network

作     者：Tian, Yingjie Sun, Shiding Tang, Jingjing 

作者机构：Univ Chinese Acad Sci Sch Econ & Management Beijing 100190 Peoples R China Chinese Acad Sci Res Ctr Fictitious Econ & Data Sci Beijing 100190 Peoples R China Chinese Acad Sci Key Lab Big Data Min & Knowledge Management Beijing 100190 Peoples R China Univ Chinese Acad Sci Sch Math Sci Beijing 100049 Peoples R China Southwestern Univ Finance & Econ Fac Business Adm Sch Business Adm Chengdu 611130 Peoples R China Southwestern Univ Finance & Econ Inst Big Data Chengdu 611130 Peoples R China 

出 版 物：《NEURAL NETWORKS》 (神经网络)

年 卷 期：2022年第146卷第0期

页      面：69-84页

核心收录：

学科分类：1002[医学-临床医学] 1001[医学-基础医学(可授医学、理学学位)] 0812[工学-计算机科学与技术（可授工学、理学学位）] 10[医学] 

基　　金：Na-tional Natural Science Foundation of China [12071458  71731009  71901179  71991472] 

主　　题：Multi-view learning Information fusion Teacher-student Network Knowledge distillation 

摘      要：Multi-view learning aims to fully exploit the view-consistency and view-discrepancy for performance improvement. Knowledge Distillation (KD), characterized by the so-called Teacher-Student (T-S) learning framework, can transfer information learned from one model to another. Inspired by knowledge distillation, we propose a Multi-view Teacher-Student Network (MTS-Net), which combines knowledge distillation and multi-view learning into a unified framework. We first redefine the teacher and student for the multi-view case. Then the MTS-Net is built by optimizing both the view classification loss and the knowledge distillation loss in an end-to-end training manner. We further extend MTS-Net to image recognition tasks and present a multi-view Teacher-Student framework with convolutional neural networks called MTSCNN. To the best of our knowledge, MTS-Net and MTSCNN bring a new insight to extend the Teacher-Student framework to tackle the multi-view learning problem. We theoretically verify the mechanism of MTS-Net and MTSCNN and comprehensive experiments demonstrate the effectiveness of the proposed methods. (C) 2021 Elsevier Ltd. All rights reserved.