Welcome to Journal of University of Chinese Academy of Sciences,Today is

Journal of University of Chinese Academy of Sciences ›› 2025, Vol. 42 ›› Issue (2): 268-275.DOI: 10.7523/j.ucas.2023.071

• Research Articles • Previous Articles    

Solutions of cross-entropy loss with spectral decoupling regularization

HU Yinhan, GUO Tiande, HAN Congying   

  1. School of Mathematical Sciences, University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2022-12-29 Revised:2023-09-01

Abstract: In this paper, we study the effect of spectral decoupling with different strengths on over-parameterized models. In the absence of weight decay, we show that the models obtained by spectral decoupling of different strengths are equivalent. When there is a small weight decay, we use the second-order Taylor expansion of the objective function to obtain an approximate solution. Analyzing the approximate solution, we find that reducing the spectral decoupling has the effect of enhancing the weight decay, which is directly equivalent in the binary classification problem. Finally, we verify our analytical conclusions through experiments.

Key words: cross-entropy loss, spectral decoupling regularization, weight decay, gradient starvation, neural network

CLC Number: