Associative Computational Model of Mirror Neurons that connects Missing Link between Behaviors and Symbols Tetsunari Inamura *1.2 Yoshihiko Nakamura .2.1 Moriaki Shimozaki .2 • 1 Japan Science and Technology Corporation, CREST program, JAPAN • 2 Dept. of Mechano-Informatics, School of Eng., The University of Tokyo, JAPAN Abstract Behavior recognition process and behavior genera- tion process have a close relationship in humans' brains. It is expected that humans' brains under- stand the meaning of behavior and create symbols through co-development of recognition and genera- tion processes. In this paper, we propose a novel method for the integration of behavior patterns and symbols using associative memory in order to real- ize the co-development processing. In the model, be- havior recognition process and generation process are practiced based on a mutual dynamics. We also con- firmed the feasibility of the method on humanoid sim- ulator. 1 Introduction The research of humanoid robots has a long history and has accumulated a substantial amount of lit- erature. The focus of early efforts was mostly on the dynamics, motion planning, and control of biped walk. Although it has not yet reached the level of complete solution with full of liability and adapt- ability, the hardware technology has been established for building the autonomous humanoids. The focus of humanoid research is now about to extend to the research on human-like intelligence. The mirror neurons[I] are found in the frontal lobes of human and primate. They activate themselves not only when he/she observes a specific behavior of the others, but also when he/she intends to act the same behavior. Furthermore, the mirror neurons are located at neither motor field nor sensory field but broker's field which has close relationship with language field. It implies that the behavior percep- tion process and behavior generation process might be integrated as an organization which has a close relationship between symbol manipulation. In the field of cognitive science, a hypothesis of mimesis[2] also drawing attentions. Mimesis is the primitive skill of communicative intelligence with im- itation learning; understanding the others' behav- iors and constructing self-behaviors. Especially, the primates who cannot manipulate speech languages can make social communications through behavior imitation[3]. On the other hand, Deacon[4] advo- cates a hypothesis that the brains of humans have co- evolved with symbol communication, in other words, humans' high-degree intelligence cannot be realized without skill of symbol manipulation. As a conse- quence, a suggestion is arises that the origin of hu- man intelligence results from the skill of imitation learning which is the strong combination of behavior perception and generation. We believe that the theory of integration between be- havior perception and generation leads to the break- through for the synthesis theory of artificial intel- ligence, like an embodiment of humanoids, symbol grounding problems, and so on. Although many hu- manoid researches treated the relation of imitation learning and intelligence [5] [6] [7] [8], few arguments were made on the connection of behavior cognition and behavior performance. We have proposed an in- tegration model for behavior perception and gener- ation using Hidden Markov Models[9], however, the mathematical background of the system has a great gulf between the concept of mirror neurons. The goal of this paper is to provide a mathematical framework of mimesis as a computational model of mirror neu- rons, based on associative memory using recurrent neural networks. In section 2, we describe the advantages and issues of time series data recognition and generation based on associative memory. In section 3, we propose a novel extension method for the associative memory which enables the system to memorize much more data and to decrease the calculation time. In section 4, we explain the mechanism of memorizing, gener- ation, and recognition. In Section 5, experiment on the humanoid simulator is shown, and discussing the result in section 6. Proceedings of the 2002 IEEE/RSJ Intl. Conference on Intelligent Robots and Systems EPFL, Lausanne, Switzerland • October 2002 0-7803-7398-7/02/$17.00 ©2002 IEEE 1032