Resolution of Singularities Introduced by Hierarchical Structure in Deep Neural Networks.

Abstract

We present a theoretical analysis of singular points of artificial deep neural networks, resulting in providing deep neural network models having no critical points introduced by a hierarchical structure. It is considered that such deep neural network models have good nature for gradient-based optimization. First, we show that there exist a large number of critical points introduced by a hierarchical structure in deep neural networks as straight lines, depending on the number of hidden layers and the number of hidden neurons. Second, we derive a sufficient condition for deep neural networks having no critical points introduced by a hierarchical structure, which can be applied to general deep neural networks. It is also shown that the existence of critical points introduced by a hierarchical structure is determined by the rank and the regularity of weight matrices for a specific class of deep neural networks. Finally, two kinds of implementation methods of the sufficient conditions to have no critical points are provided. One is a learning algorithm that can avoid critical points introduced by the hierarchical structure during learning (called avoidant learning algorithm). The other is a neural network that does not have some critical points introduced by the hierarchical structure as an inherent property (called avoidant neural network).