Training Behavior of Deep Neural Network in Frequency Domain

• Published in: Neural Information Processing Vol. 11953; pp. 264 - 274
• Main Authors: Xu, Zhi-Qin John, Zhang, Yaoyu, Xiao, Yanyang
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2019; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Deep learning; Deep Neural Network; Fourier analysis; Generalization
• DOI: 10.1007/978-3-030-36708-4_22
• ISBN: 303036707X; 9783030367077
• ISSN: 0302-9743

Why deep neural networks (DNNs) capable of overfitting often generalize well in practice is a mystery [24]. To find a potential mechanism, we focus on the study of implicit biases underlying the training process of DNNs. In this work, for both real and synthetic datasets, we empirically find that a DNN with common settings first quickly captures the dominant low-frequency components, and then relatively slowly captures the high-frequency ones. We call this phenomenon Frequency Principle (F-Principle). The F-Principle can be observed over DNNs of various structures, activation functions, and training algorithms in our experiments. We also illustrate how the F-Principle helps understand the effect of early-stopping as well as the generalization of DNNs. This F-Principle potentially provides insight into a general principle underlying DNN optimization and generalization.