Multimodal Representation of Hand Grasping based on Deep Belief Nets

摘要

In human brain, different sensor information is thought to be processed in different area and integrated in parietal area. Fig. 1 shows a model of neural mechanism for grasping proposed by Oztop et al (Oztop et al., 2006). As shown in this gore, the information of hand and object is processed separately and important features for grasping are extracted in the hierarchical network. In this paper, we aim to construct a hierarchical model for grasping like brain model. The hierarchical model is thought to be plausible as developmental model, because an infant learns its grasping skills gradually in the developing process (Case-Smith and Pehoski, 1992). For this purpose, we adopt deep belief network (DBN), proposed by Hinton, for representing the multimodal information in grasping, in which one modal information is self organized to extract statistical information of given data and different modal information are easily integrated in the hierarchical architecture (Hinton, 2007).