Research on Capsule Network Optimization Structure by Variable Route Planning

Abstract

Convolutional Neural Networks (CNN) perform quite good on image classification and object recognition tasks. To solve the problem that CNN is not robust to affine transformation and can not consider the spatial relationship between objects in images, Capsule Network (CapsNet) is proposed and the best accuracy of MNIST dataset has achieved. It shows developability on other datasets and its performance is significantly higher than CNN. However, the data categories in the existing routing algorithms of the CapsNet need to be given in advance, which is very difficult to estimate. In a complex dataset or a small sample dataset, calculating the maximum likelihood solution is more complicated, and it will lead to an inevitable singularity problem. For these shortcomings of the routing algorithm, this paper first proposes a variational routing algorithm. The algorithm does not bring additional computational burden, can automatically determine the most suitable data category, and can effectively avoid over-fitting problems. The algorithm uses the variational distribution to integrate the prior distribution information to increase the regularization limit, maximize the lower bound of the model evidence, and avoid the singularity problem caused by computing the maximum likelihood function. In this paper, the MNIST dataset is classified by a simple CapsNet architecture, and the result is encouraging.