An Optimization view on Squash Function of CapsNet

Abstract

In CapsNet, a bounded measure of the modulus length of the feature is needed, so Squash function is used to compress the feature vector. This paper discusses the definition of Squash function, redefines Squash function based on the idea of information gain rate of decision tree, and constructs CapsNet model on this function. By testing on MNIST, Fashion-MNIST and Cifar-10 datasets, the experimental results show that the Squash function defined in this paper has better classification performance than the traditional Squash function in CapsNet model.