On the Expected Complexity of Maxout Networks

Abstract

Learning with neural networks relies on the complexity of the representable functions, but more importantly, the particular assignment of typical parameters to functions of different complexity. Taking the number of activation regions as a complexity measure, recent works have shown that the practical complexity of deep ReLU networks is often far from the theoretical maximum. In this work, we show that this phenomenon also occurs in networks with maxout (multi-argument) activation functions and when considering the decision boundaries in classification tasks. We also show that the parameter space has a multitude of full-dimensional regions with widely different complexity, and obtain nontrivial lower bounds on the expected complexity. Finally, we investigate different parameter initialization procedures and show that they can increase the speed of convergence in training.

Published at NeurIPS 2021, 47 pages, 18 figures

References in corpus (7)

• PyTorch: An Imperative Style, High-Performance Deep Learning Library
• Complexity of Linear Regions in Deep Networks
• Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
• On the Number of Linear Regions of Convolutional Neural Networks
• Deep ReLU Networks Preserve Expected Length
• Sharp bounds for the number of regions of maxout networks and vertices of Minkowski sums
• Using activation histograms to bound the number of affine regions in ReLU feed-forward neural networks