Abstract

We investigate the combinatorics of max-pooling layers, which are functionsthat downsample input arrays by taking the maximum over shifted windows ofinput coordinates, and which are commonly used in convolutional neuralnetworks. We obtain results on the number of linearity regions of thesefunctions by equivalently counting the number of vertices of certain Minkowskisums of simplices. We characterize the faces of such polytopes and obtaingenerating functions and closed formulas for the number of vertices and facetsin a 1D max-pooling layer depending on the size of the pooling windows andstride, and for the number of vertices in a special case of 2D max-pooling.

 

Quick Read (beta)