Connecting Brains and Robots by Computational Theories

In ATR Computational Neuroscience Laboratories, a series of computational, imaging, neurophysiological and robotics studies explored several key concepts such as cerebellar internal models, multiple internal models, MOSAIC, imitation learning, biologically motivated robot biped locomotion, modular and hierarchical reinforcement learning models. Recent efforts in ATR CNS labs including computational-model based imaging, hierarchical variational Bayesian method in fMRI-MEG combination, non-invasive decoding of -neural representations, and robotics experiments could be the bases of the new methodology in neuroscience. Suppose you have a computational theory, which postulates that some brain networks solve some computational problems and a specific brain locus contains a specific computational representation. You extract this information either by some non-invasive method or unit recording, and manipulate this by altered computational algorithms derived from the theory. The altered or processed information is fed back into a robot and then to the brain by appropriate methods (e.g. visual or tactile feedbacks, TMS, electrical stimulation). I explain bases of this new approach.