Fast convergent capsule network with applications in MNIST

Abstract

Capsule network is a new neural network architecture, which avoids the problem of location information loss due to the pool operation of the convolution neural network. The capsule network uses vector as input and output and dynamic routing updates parameters, which has better effect than convolution neural network. In this paper, a new activation function is proposed for the capsule network and the least weight loss is added to the loss function. The experiment shows that the improved capsule network improves the convergence speed of the network, increases the generalization ability, and makes the network more efficient.