BIMSA >
控制理论和非线性滤波讨论班
Deep Network Approximation: Achieving Arbitrary Accuracy with Fixed Number of Neurons
Deep Network Approximation: Achieving Arbitrary Accuracy with Fixed Number of Neurons
组织者
演讲者
孙泽钜
时间
2023年08月07日 14:30 至 15:00
地点
数学系理科楼A-203
摘要
In this talk, we will review the paper concerning the approximation capability of deep neural network by Z. Shen et al. In this paper, the authors theoretically proved that deep neural networks with fixed number of neurons have the capability of approximating continuous functions in compact domains with arbitrary precision. Different from the popular ReLU network, the activation function is a combination of a triangular-wave function and the soft sign function. It can be proved that the parameters in such neural networks can be chosen to grow only polynomially with respect to the dimension.