From throwing a baseball to playing the piano to typing on keyboards, human beings are constantly learning new sensorimotor skills. During learning, synaptic connections in the brain must be modified to form a motor memory. Further, this modification seems both permanent and robust: a sensorimotor skill, once learned, tends to persist throughout the course of a lifetime regardless of its salience (recall the old adage of never forgetting how to ride a bike). Despite the importance of motor memories, their distinctive features, and their ubiquity in vertebrate behavior, little is known about the computational principles and mechanisms that subserve the acquisition of sensorimotor skills. This US-Canadian collaborative project takes an interdisciplinary approach aimed at elucidating neural mechanisms of motor memory formation and unifying -- under a common theoretical principle -- the findings of single-neuron recording studies with established behavioral results. The theory that is proposed makes the following testable prediction: as the level of behavioral expertise in a specific task increases, the neural representation for that skill becomes more selective. By selective, it is meant that a neuron significantly recruited during the performance of the skill tends, with practice, to specialize by firing only when that skill is performed (and not when related skills are performed).
Central to the theory is a geometric interpretation of "biologically plausible" sensorimotor neural networks, in which neurons are modeled as noisy signal processors and synaptic change is modeled as a noisy morphological process. Because of the high noise levels, it is shown that the system must be "hyperplastic" -- that is, the learning rate must be unusually high in order to compensate for the noise and operate at an acceptable performance level. Geometrically, the solution for a skill can be represented as a manifold in the weight space of the network. To learn multiple skills, a network configuration must be attained such that the solution manifolds intersect. To learn multiple skills without noise leading to destructive interference, the network must arrive at a point where the intersecting solution manifolds are orthogonal. With this principle of orthogonality, the neurophysiological predictions described above can be explicitly formulated. These predictions will be tested with an experimental method -- involving floating microelectrode arrays and antidromic stimulation -- that enables the identifiably same neuron to be recorded from for multiple days/weeks, while a behaving animal learns a task. Finally, psychophysical predictions of the theory will also be tested.
This project is jointly funded by Collaborative Research in Computational Neuroscience and the OISE Americas program. A companion project is being funded by the Canadian Institutes of Health Research.
