Neural network models of motor timing and coordination

摘要

Purposive motor behavior in an animal or robot requires simultaneous coordination of many degrees of freedom in time and space. Due to continual changes in actor and environment, such coordinative functions are adaptive. This dissertation describes how networks of neurons provide adaptive motor timing in response to predictive stimuli, and adaptive coordination of multiple degrees of freedom during visually-guided reaching. The cerebellum adaptively times motor responses such that conditioned responses (CRs) match external constraints like the delivery of aversive unconditioned stimuli (US). A neural model of this brain circuit reproduces key physiological and behavioral features of the conditioned nictitating membrane response in rabbits, such as: The observed CR topography at the nucleus interpositus (NI) with the NI response preceding and the CR peak amplitude occurring at the expected US onset time. During training, CR onset latency decreases, CR amplitude increases, and climbing fiber activity decreases. Maximal conditioning occurs at interstimulus intervals (ISIs) of 200-400ms. The CR peak tracks ISI changes, and mixed training at two ISIs produces a double-peaked CR. The cerebellar timing model suggests that Purkinje neuron hyperpolarizing responses are timed by the metabotropic glutamate receptor second messenger system. Calcium dependency of the inositol trisphosphate (IP$\sb3$) receptor in the model generates characteristic responses observed in phosphoinositide response systems, as in invertebrate photoreceptors. Slow IP$\sb3$ generation yields a delayed intracellular calcium signal which drives a calcium-dependent potassium conductance and releases NI from inhibition, allowing a timed CR for ISIs of 80-4000ms. Conditioning occurs through the phosphorylation/dephosphorylation processes of long-term depression (LTD) and potentiation (LTP) at parallel fiber-Purkinje synapses. The straight lines of primate visually-guided reaches are learned behaviors. Straight line trajectories require the brain to transform the visually perceived line into a nonlinear set of motor commands. An adaptive network for learning the transformation from visual space velocities to joint space velocities is obtained by a gradient descent learning law. The learned mapping is a singularity-robust pseudoinverse of the arm Jacobian. A reaching model including this network reproduces primate/human straight-line hand trajectories with bell-shaped velocity profiles.