Deep multi-view learning methods: A review

Abstract

Multi-view learning (MVL) has attracted increasing attention and achieved great practical success by exploiting complementary information of multiple features or modalities. Recently, due to the remarkable performance of deep models, deep MVL has been adopted in many domains, such as machine learning, artificial intelligence and computer vision. This paper presents a comprehensive review on deep MVL from the following two perspectives: MVL methods in deep learning scope and deep MVL extensions of traditional methods. Specifically, we first review the representative MVL methods in the scope of deep learning, such as multi-view auto-encoder, conventional neural networks and deep brief networks. Then, we investigate the advancements of the MVL mechanism when traditional learning methods meet deep learning models, such as deep multi-view canonical correlation analysis, matrix factorization and information bottleneck. Moreover, we also summarize the main applications, widely-used datasets and performance comparison in the domain of deep MVL. Finally, we attempt to identify some open challenges to inform future research directions.