Javier Gonzalez Anaya (Harvey Mudd College)

This is the continuation of the semester’s joint seminar with the Universidad Nacional de Colombia-Manizales.

Title: Enumerating linearity regions of max-pooling layers in convolutional neural networks

Abstract: Convolutional neural networks (CNN’s) are central tools in the application of machine learning to text, audio and image processing. Their success stems from the ability of these networks to identify key features in complex datasets at a relatively low computational cost. Max-pooling layers (MPL’s) are key components of CNN’s that reduce the number of parameters used by the network while making it more robust to small changes in the input data. From a mathematical point of view, MPLs are piecewise-linear functions, and their number of linearity regions can be interpreted as a measure of complexity of the layer. In this talk I will explain how we can use combinatorial techniques to count these linearity regions, and survey our current results in the area.