A Linear Frequency Principle Model to Understand the Absence of Overfitting in Neural Networks

• 作者: Yaoyu Zhang Tao Luo Zheng Ma Zhi-Qin John Xu
• 来源期刊: 中国物理快报（英文版）
• 年，卷（期）: 2021,（3）
• 所属期刊栏目: CROSS-DISCIPLINARY PHYSICS AND RELATED AREAS OF SCIENCE AND TECHNOLOGY
• 页码范围: 121-126
• 页数: 6页
• 语种: 英文
• DOI: 10.1088/0256-307X/38/3/038701

摘要

Why heavily parameterized neural networks (NNs) do not overfit the data is an important long standing open question. We propose a phenomenological model of the NN training to explain this non-overfitting puzzle.