Bayesian Network Model for Object Concept

This paper discusses a computational model for object concept formation. We propose a model of object concept based on the relationship between shape and function. Implementation of the proposed framework using Bayesian network is presented. At this point we need an explicit definition of object function. In the proposed model each function is defined as certain changes in a target object caused by the object. Therefore each function is represented by a feature vector which quantifies the changes in the target. Then the function is abstracted from these feature vectors using the Bayesian learning approach. The system can form object concept by observing the human tool use based on the abstract function and shape information. Furthermore, it is demonstrated that the learned model (object concept) enables the system to infer the property of unseen object. The system is evaluated using 35 hand tools, which reveals the validity of the proposed framework.