Date of Publication :17th November 2017

Abstract: The Convolutional Neural Network is a powerful field of study that has been the go-to solution to analyze and classify visual imagery in the recent phase.However, the current state-of-the-art CNN's are facing a reliability crisis when the images are rotated or tilted. There is a drastic fall in their performance when an image isn’t close to the ones in the training data-set. A change in orientation and spatial positioning of features, also, has a negative impact on its working. In this paper, we seek answers to these problems by using Capsule Networks, a recent advancement in deep learning. Instead of adding layers, the Capsule Networks work on nesting the layers of a neural network. These nested layers are called capsules. This increases the internal structure of each neuron rather than the depth of the network. Such a structure helps to improve the learning process of machines and renders them close to human intuition. In our work, we explain Capsule networks for Image classification systems and throw light on their robustness.

Reference :