RED-Nets: Redistribution Networks for Multi-View Classification

Abstract Multi-View Learning (MVL) focuses on effectively utilizing information from different views to promote analysis performance. To explore the correlation among multiple views, most existing multi-view algorithms usually maximize the correlation between different views for consistency. However, there are two main limitations for this strategy. First, the information from different views cannot be fully integrated only using some specific constraints on pair of views. Second, beyond consistency, the diversity can basically explore the underlying complementary information among multiple views. Therefore, we propose a novel multi-view classification model termed Redistribution Networks for Multi-View Classification (RED-Nets), which is able to jointly explore consistency and diversity in a flexible manner. Specifically, in our model, different (original) views are integrated first and then redistributed into multiple pseudo views to simultaneously capture the consistency and complementarity. The redistributed way is endowed with the ability of searching for the optimal combination for consistent classification and high complementarity from the original views, breaking the barriers among these views. Experimental results on multiple benchmark datasets validate that the proposed RED-Nets is more effective than the state-of-the-art methods.