The hebbian theory of sensorimotor learning and its application to birdsong learning
Institute of Neuroinformatics, ETH and University of Zurich
How can distributed sensorimotor circuits, using only local synaptic plasticity mechanisms, enable organisms to rapidly learn complicated motor sequences? For example, how can a songbird, exposed only to the sensory experience of a tutor's song, learn to produce complex sequences of muscle activations that generate the same song? I introduce a new paradigm for birdsong learning. First, in a sensory phase, tutor song exposure creates a template circuit that can replay an auditory representation of the tutor song. Second, as the bird begins to babble, it learns an inverse model that maps auditory song representations into appropriate motor patterns required to generate song. Fourth, the bird feeds the output of the sensory template into the inverse model, to generate song and further train motor circuits. We show these phases can all be implemented through local hebbian plasticity rules. Overall, this new paradigm not only explains one-shot learning, obtained by feeding a rapidly learned sensory representation into the inverse model, but also makes several testable behavioral and physiological predictions about the development of the song circuit.