Optimal Match of Sample Size and Data Dimension for Benign Overfitting
数学学科创建110周年系列报告
报告题目(Title)：Optimal Match of Sample Size and Data Dimension for Benign Overfitting
报告人(Speaker)：郑术蓉 （东北师范大学）
地点(Place)：后主楼1220
时间(Time)：2025年5月21日 周三9:00-10:00
邀请人(Inviter)：段玉萍
报告摘要
Studying the asymptotics of generalization errors is an important and practical topic in large models. The existing literature is mainly based on the limits of the generalization errors of linear models, random feature regression or kernel ridge regression under different convergence regimes. Though some of them derived the double descent curves, they didn't give the optimal match of the number of model parameter and the number of samples. The optimal match can tell us that giving the sample size, how large a model can be trained to control the generalization error. This is useful for guiding the training of large models. Moreover, this paper also theoretically derived that overfitting model is noise-resistant, double descent phenomenon may disappear and the prediction accuracy may be improved by introducing the optimal tuning parameter. Simulation studies lend further support to our proposed method and show that our theoretical results are appropriate.
