Seminar| Institute of Mathematical Sciences
Time: Tuesday, December 19th, 2023 , 13:30-14:30
Location：RS518, IMS
Speaker: Zhiqin Xu, Shanghai JiaoTong University
Abstract: Why do neural network models that look so complex usually generalize well? To understand this problem, we study deep learning training and find that some simple implicit regularization effects. We focus on the condensation phenomenon,which indicates neurons bias towards the same during the training.This talk will discuss the mechanism and potential application of condensation.