Kawato Dynamic Brain Project
Japan Science and Technology Corporation
October 1996 - September 2001


It has become widely accepted that in order to elucidate the various functions of the brain, three levels of research are essential: the lowest level of hardware (molecules, synapses, and neural circuits), the middle level of algorithms and representations, and the highest level of computational theory. Unfortunately, computational theory has achieved limited success: although the three levels impose strong constraints among each other, past research has focused primarily on isolated studies within the highest level. Also, despite the well-known strong reciprocal influence between different brain functions, such as vision and movement control, research has tended to concentrate on only one function, without coherently studying the overall information processing from inputs to outputs.

Recently, based on experimental evidence from visuo-motor coordination tasks, Mitsuo Kawato has proposed a bi-directional theory of cognition and motor control. This theory is built on the dynamical interaction among models in the brain that represent parts of the external world. The validity of this theory has been tested in biological studies of eye-movement control, where, for instance, it was shown that a particular part of the cerebellum constitutes an inverse-dynamics model of the eye dynamics. In a variety of manipulation tasks, it was furthermore demonstrated that this theory can be used for the learning and control in robots.

Research Strategy

The Kawato Dynamic Brain Project is grounded in the study of sensory-motor integration. Special emphasis is being placed on a coherent investigation from inputs to outputs. Emphasized are simultaneous studies at the hardware, algorithmic, and computational levels. Hopefully it will also be possible to add to the understanding of the mechanism of human cognitive functions, such as thinking, emotion and communication, as these higher functions are built on top of the brain's ability to utilize models in a dynamic fashion. Three major areas of the research are being pursued: computational neurobiology, computational psychology, and computational learning. All them are rooted in the study of sensory-motor integration and closely linked together.

1. Regarding computational neurobiology, an attempt is being made to bridge the gaps between computational theories and neurobiological reality in the form of biologically plausible network models. The models can be of different physical levels, from macroscopic ones that account for the functions of different brain areas based on lesion and imaging studies, to detailed ones that take into account cellular and molecular data about the local circuitry and plasticity. These models are not only testable experimentally but provide novel experimental designs crucial for understanding the brain's computational mechanisms. The research subjects include: 1) learning and control of sequential action through reinforcement (basal ganglia, cerebellum, cortical motor areas); 2) functions of local circuit, intra cellular dynamics, and neuromodulators in motor learning (cerebellum, striatum); 3) representations of complex temporal patterns in the brain for perception and motion generation (in songbirds and humans); 4) models of emotion and consciousness as the global regulatory system of the brain.
2. Regarding computational psychology, sensory-motor control research has recently come to provide a comprehensive account of how the central nervous system (CNS) may plan purposeful acts starting from sensory input to motor output. Thus, computational theories and models concerning human motor control and learning are being examined using behavioral (psychological) experiments and non-invasive methods of investigating human brain activities (e.g. fMRI, PET and MEG). Hopefully, it will become possible: 1) to identify how, where, and when computational problems of motor control and learning (e.g., trajectory planning, coordinate transformation, generation of motor commands, and acquisition of internal models) are solved in the CNS, and 2) to investigate whether human subjects can learn optimization principles in trajectory planning and whether the principles contribute to the perception of the human body motion. Also, the biological plausibility of computational models and robotic simulations proposed by the project are being examined using these methods.
3. The third research area, computational learning, is seeking to obtain an understanding of sensory-motor coordination on a system's level, particularly using statistical, physical, and mathematical modeling. The general research goals are to obtain insight into the theoretical constraints of movement generation, to device algorithms of how particular problems of motor control and learning can be solved, and to validate these algorithms by synthesizing behavior with anthropomorphic robot hardware.
Much effort is being devoted to comparisons of synthesized behavior with human behavior in collaboration with the computational psychology approach, as well as joint studies of functional models of brain data with computational neurobiology. Research projects include topics of statistical learning, reinforcement learning, and neural network learning for problems of sensory-motor control, topics of dynamical systems theory for motor pattern generation, and the study of learning from demonstration in light of movement primitives, movement segmentation, and the sequencing of complex motor acts.
It is believed that if the function of the brain can be understood, it should become possible to implement it in a robot, or any other form of artificial machine. Thus, constructing and programming humanoid robots that can incorporate the computational models and algorithms developed to explain learning in sensory-motor integration is being actively pursued.

Last modified: Fri Nov 22 19:19:50 JST 1996