Active Exploration for Unsupervised Object Categorization Based on Multimodal Hierarchical Dirichlet Process

Abstract

This paper describes an effective active exploration method for multimodal object categorization using a multimodal hierarchical Dirichlet process (MHDP). MHDP is a type of multimodal latent variable models, e.g., multimodal latent Dirichlet allocation and multimodal variational autoencoder, that enables a robot to perform unsupervised multimodal object categorization on the basis of different types of sensor information. The goal of the active exploration is to reduce the number of actions executed to collect multimodal sensor information from a variety of objects to acquire knowledge on object categories. The active exploration method employing the information gain (IG) criterion for MHDP is described by extending the IG-based active perception method. Exploiting the submodular property of IG in MHDP, greedy and lazy greedy algorithms with a certain theoretical guarantee of performance are proposed. The effectiveness of the proposed method is evaluated in a robot experiment. Results show that the proposed active exploration method with the greedy algorithm works well, and it significantly reduces the step for exploration. Further, the performance of the lazy greedy algorithm is found to deteriorate at times, due to the estimation error in the IG, differently from that of active perception.