Discriminative Deep Generalized Dependency Analysis for Multi-View Data

In recent years, a surging interest is noted for combining the information of multiple views to obtain a joint representation of the given data. In multi-view data analysis, the joint representation should be learned from the given input views in such a way that the view-specific information as well as the cross-view dependency are preserved properly. In the context of cross-view dependency, it is expected that both view-consistency and view-discrepancy are addressed simultaneously. Discriminability of the joint representation is also an important aspect in the classification problem. In this regard, a novel deep learning model is proposed to efficiently encapsulate the underlying data distribution over the space of input views. Considering both consensus and complementary principles, a loss function is introduced, based on the concept of the Hilbert–Schmidt independence criterion, to capture the relevant cross-view information from the given multi-view data. Incorporating the supervised information of sample categories not only enhances the discriminative ability of the model but also allows it to classify the given samples into different categories. An upper bound on the error probability of the proposed deep model is estimated in terms of the model architecture. It facilitates determining the optimal architecture of the proposed model for each database. The proficiency of the model is studied on numerous application domains with reference to several state-of-the-art multi-view classification algorithms.