Category formation of real objects using Multimodal Variational Autoencoder

Abstract

We propose a neural network-based unsupervised object categorization method for a robot using multimodal sensor information. The method is an extension of Multimodal Variational Autoencoder (MVAE). In the proposed method, Dirichlet prior is introduced for giving MVAE a clustering capability in the same way as Multimodal latent Dirichlet allocation (MLDA) that has been used for multimodal object categorization by a robot. We performed comparative experiments with MLDA using both real objects and synthetic data. The results show that our proposed model has a reduced computational costs compared to MLDA without deteriorating the classification accuracy.