An Experimental Study on Capsule Networks

In this work we perform experiments with the recently published work on Capsule Networks. Capsule Networks have been shown to deliver state of the art performance for MNIST and claim to have greater discriminative power than Convolutional Neural Networks for special tasks, such as recognizing overlapping digits. The authors of Capsule Networks have evaluated datasets with low number of categories, viz. MNIST, CIFAR-10, SVHN among others. We evaluate capsule networks on two datasets viz. Traffic Signals, Food101, and CIFAR10 with less number of iterations, making changes to the architecture to account for RGB images. Traditional techniques like dropout, batch normalization were applied to capsule networks for performance evaluation.