Multi-View Variational Autoencoder for Robust Classification against Irrelevant Data

Multi-view variational autoencoders (MVAEs) can extract latent variables with high discriminative power for classification by considering correlations among multi-view data, i.e., multiple kinds of data. However, if we input irrelevant data, i.e., multiple kinds of data that capture the different object, to MVAEs, discriminative power is reduced. To solve this problem, we propose an MVAE including a novel objective function. Our proposed MVAE reconstructs multi-view data without the negative effect of irrelevant data. Specifically, we derive an objective function that focuses on latent variables of relevant data, i.e., multiple kinds of data that capture the same object. Experimental results show that the proposed method improved the discriminative power of latent variables even if irrelevant data are input.