Towards manipulative neuroscience based on brain-network -interface

Abstract

Event Abstract Back to Event Towards manipulative neuroscience based on brain-network -interface Mitsuo Kawato1* 1 ATR Computational Neuroscience Labs, Japan The cerebellar internal model theory postulates that the cerebellar cortex acquires many internal models of controlled objects, dynamical processes in the external world dependent on long-term depression (LTD) of Purkinje cells. It predicts that the climbing fiber inputs to Purkinje cells carry the feedback motor command and can supervise learning of inverse dynamics models. Many experimental supports were obtained from the ventral paraflocculus of the cerebellum during monkey control of ocular following responses. fMRI studies mapped forward and inverse models of manipulated objects and tools in the cerebellar cortex. Kinetic models of LTD [1,2] suggest a cascade of excitable and bistable dynamical processes, which may resolve plasticity-stability dilemma at single spine level. That is, even a single pulse of climbing fiber input combined with an early train of several parallel fiber pulses can induce Ca2+ induced Ca2+ release via IP3 receptors on ER. The MAPK positive feedback loop leaky integrates resulting large Ca2+ elevation and if it crosses the threshold then the state moves to the depressed equilibrium. These models explain diverse LTD experiments and clearly demonstrate that LTD is a supervised learning rule, and not anti-Hebbian as erroneously characterized. The MAPK positive feedback loop model [1] was recently supported by a Ca2+ photo-uncaging and imaging experiment [3] that supports LTD all-or-none character. In [3], the most important information within the system, Ca2+ can be measured and manipulated directly, thus the theory was quite rigorously proved. At the system level, the cerebellar internal models for arm movements are still debated and we need to develop a method to directly manipulate information. In our "Understanding brain by Creating Brain"-approach, high-performance humanoid robots have been developed, and sensory-motor coordination problems were investigated. We explored several computational theories such as MOSAIC, imitation learning, biologically motivated robot biped locomotion, modular and hierarchical reinforcement learning models. Our paradigms include computational-model based imaging and non-invasive decoding of neural representations. However, we got frustrated that most experiments can show just temporal correlation between data and theory but not causality, unlike the above spine-level LTD story. We hope that manipulative neuroscience based on real time feedback of decoded neural information could be a resolution to this difficulty. Hierarchical Bayesian approach to combine fMRI and MEG, or NIRS and EEG, or neural decoding with machine learning techniques, and real-time control of robots with decoded information are expected to be key technological elements. Demonstrations of real-time robot hand control by decoded information from fMRI (http://www.atr.jp/html/topics/press_060526_e.html) in collaboration with Honda, control of a humanoid robot locomotion-like movements from monkey brain in collaboration with Miguel Nicolelis [4], prediction of wrist joint angular acceleration from the primary motor cortex currents estimated by hierarchical Bayesian MEG, or estimation of visual target velocity from MST currents, etc could be technical bases for this new approach.