Abstract
Incremental learning of knowledge and context dependency of recognition are important characteristics for an intelligent machine in the real world environment. Unknown objects may appear among known objects in such environment and the context requires change of the recognition result even if the input is the same. The system has to learn which object is unknown, what knowledge is necessary and how the context acts on the recognition process. Associative memory model PATON (Pattern+ton) has been proposed to realize such context dependency of recognition on the basis of the attention vectors. External inputs and candidates for the recognition can both be selected by an attention vector. In this paper, we propose a method of incremental learning involving the attention vectors and the knowledge, based on a reward through conversational interactions between PATON and the environment.
