Improving Training time in Capsule Neural Network

Abstract

Chiadika Electrical & Computer Engr Mathematics department Electrical & Computer Engr Brunel University, Seoul National University Brunel University, London, United Kingded the Attention Routing CapsuleNet (AR CapsNet) as proposed by Jaewoong Choi et al. in Attention Routing between Capsules. The AR-CapsNet is an enhanced version of CapsNet which, uses a new and different routing and activation function. The unique routing style is Attention routing which, is simply capsules been routed, with the help an attention module and a fast-forward pass but, what is most important is that the spatial information is kept, which is the primary reason behind Capsules. Primarily, the in-built interpretation of the dynamic routing is finding a focal point of the prediction capsules. As well known, emphasis on preserving a vector orientation is what activation functions and its variant deal majorly on; the activation function used is capsule activation because it focuses is on how a capsule-scale activation function performs. The model was trained on the MNIST and CIFAR-10 datasets and classification tasked against AR CapsNet and CapsuleNet. The model showed almost accuracy with less training parameters and less training time.