Representative Primary Capsule in Capsule Network Architecture for Fast Convergence
Eashan Dash, J Mercy Faustina, B. Sivaselvan · 2020
Deep neural networks like the Convolutional Neural Networks (CNN) is a popular state of the art method for various computer vision applications used for image classification and detection. However, they do not consider spatial relationships between various objects in the image. The Capsule Network (CapsNet) architecture proposed by Geoffrey Hinton consists of capsules instead of neurons. The capsule network is developed to overcome the drawbacks of CNN such as back-propagation, translation variance and pooling layers. This paper presents a comparison between the CNN and capsule network with respect to accuracy and training time to identify suitable applications which can be solved using capsule network. For smaller datasets, capsule network perform better than CNN. For this reason, the capsule network is implemented for safety critical system like autonomous vehicles. In order to train a fully autonomous vehicle, huge dataset covering all possible road driving scenario is required. Unavailability of such a huge dataset makes capsule network an apt choice for solving the problem. This paper finally proposes a new design for the capsule network which helps the model to achieve faster convergence leading to reduced training time for all kinds of datasets. The enhanced dynamic routing agreement based capsule network model uses a new representative primary capsule in place of original primary capsules for fast convergence. The proposed model achieves 2.8 times faster convergence on the MNIST dataset.