Reinforcement Learning of Bird Vocalizations Using a Discrete Motor Space

Vocal learning in songbirds provides a model for studying the acquisition of complex motor behaviors. In particular, the zebra finch is one of the best-characterized models, acquiring its song during development through comparison of its vocalizations with those of a tutor. We propose a reinforcement learning algorithm, inspired by Doya and Sejnowski, to reconstruct experimentally recorded respiratory pressure syllables associated with song production. The algorithm explores a discrete motor space in which each syllable is represented as a temporal combination of excitable base gestures activated at discrete times. These gestures are interpreted as motor manifestations of transient events in the bird’s neural dynamics. The model successfully reconstructs the analyzed syllables using a reduced number of gestures. Moreover, the learned gestures reveal the spontaneous emergence of a small set of functional primitives shared across individuals, without imposing such a repertoire during learning. These results suggest that vocal production can be represented through combinations of reusable motor primitives, providing a low-dimensional description of the motor structure underlying vocal learning. Thus, learning can be understood not as the optimization of arbitrary continuous trajectories, but as the selection and combination of a finite repertoire of motor primitives.