1: \begin{abstract}

2: Previous studies have shown that deep neural networks (DNNs) with

3: common settings often capture target functions from low to high frequency,

4: which is called Frequency Principle (F-Principle). It has also been

5: shown that F-Principle can provide an understanding to the often observed

6: good generalization ability of DNNs. However, previous studies focused

7: on the loss function of mean square error, while various loss functions

8: are used in practice. In this work, we show that the F-Principle holds

9: for a general loss function (e.g., mean square error, cross entropy,

10: etc.). In addition, DNN's F-Principle may be applied to develop numerical

11: schemes for solving various problems which would benefit from a fast

12: converging of low frequency. As an example of the potential usage

13: of F-Principle, we apply DNN in solving differential equations, in

14: which conventional methods (e.g., Jacobi method) is usually slow in

15: solving problems due to the convergence from high to low frequency. 

16: \end{abstract}

17: