Abstract

Convolutional neural networks (CNNs) are applied in many different fields such as image processing, natural language processing (NLP) and biomedical. In recent years, a new CNN architecture known as Capsule Network (Capsule-Net) has been developed to reduce some disadvantages of convolutional neural networks and improve the performance. Architecture of capsule networks is inspired by the human brain’s inverse graphics and hierarchical mapping concepts. In this study, the architecture and working of the capsule networks were examined and comparative analysis of Capsule Network and CNN on the new handwritten digits dataset called Kannada-MNIST were defined. Our experimental results show that the accuracy of capsule network model is better than the accuracy of CNNs.