One fundamental problem in deep learning is understanding the outstanding performance of deep Neural Networks (NNs) in practice. One explanation for the superiority of NNs is that they can realize a large class of complicated functions, i.e., they have powerful expressivity. The expressivity of a ReLU NN can be quantified by the maximal number of linear regions it can separate its input space into. Various results on the number of linear regions of fully-connected ReLU NNs have been obtained since 2013. However, as far as we know, there are no explicit results on the number of linear regions for Convolutional Neural Networks (CNNs) due to the lack of proper mathematical tools. In this paper, we provide several mathematical results needed for studying the linear regions of CNNs, and use them to derive the maximal and average numbers of linear regions for one-layer ReLU CNNs. Furthermore, we obtain upper and lower bounds for the number of linear regions of multi-layer ReLU CNNs. Some asymptotic results are also derived. Our results suggest that deeper CNNs have more powerful expressivity than their shallow counterparts, while CNNs have more expressivity than fully-connected NNs per parameter. To the best of our knowledge, this paper is the first work on the number of linear regions for CNNs. Various potential future directions are given at the end of this paper.
Speakers: Huan Xiong, Lei Huang, Mengyang Yu, Liu Li, Fan Zhu, Ling Shao