Multi-view Deep Gaussian Processes

Abstract

Deep Gaussian processes (DGPs) have shown their power in many tasks of machine learning. However, when they deal with multi-view data, DGPs assume the same modeling depth for different views of data, which is quite unreasonable because there are usually large diversities among different views. In this paper, we propose the model of multi-view deep Gaussian processes (MvDGPs), which takes full account of the characteristics of multi-view data. Combining the advantages of the DGPs with the multi-view learning, MvDGPs can independently determine the modeling depths for each view, which is more flexible and powerful. In contrast with the DGPs, MvDGPs support asymmetrical modeling depths for different view of data, resulting in better characterizations of the discrepancies among different views. Experimental results on multiple multi-view data sets have verified the flexibilities and effectiveness of the proposed model.