Research on Classification Performance of Small-Scale Dataset Based on Capsule Network

Abstract

At present, deep learning methods represented by artificial neural networks (ANN) has been applied in the fields of image recognition, natural language processing, intelligent control, and big data analysis. However, whether deep learning is suitable for little data has always been a controversial topic. A recent algorithm called capsule network (CapsNet) that implements an approach based on activity vectors may overcome some of the limitations of the current state of the convolution neural network (CNN) classifiers. In this study, we explore capsule networks with dynamic routing for the problem of small scale datasets. A series of experiments are conducted with capsule networks and convolution networks. Experimental results demonstrate the effectiveness of capsule networks on little data. Furthermore, when training data does not use augmentation techniques, CapsNet significantly outperform CNN, indicating that CapsNet has better generalization ability which is a very important advantage if the amount of data is relatively small.