On Multi-view Interpretation of Convolutional Neural Networks

Abstract

In this study we consider multi-view capabilities of convolutional neural networks as one of the best methods of representation learning. Multi-view learning as a machine learning technique deals with the task of learning from multiple distinct views or multiple distinct feature sets. Moreover, multi-view feature learning attempts to abstract and summarize distinct feature sets for further machine learning and pattern recognition tasks. In contrast to traditional multi-view learning methods, convolutional neural networks are able to generate representations from unstructured raw data; these features are very essential for real world applications. It is concluded that CNNs are inherently multi-view representation learning methods able to handle both natural and artificial views.