[Seminar] Intrinsic Reward in Birdsong Learning, Prof. Hahnloser
Birdsong development depends on a sensory phase of template acquisition and a sensorimotor phase of trial-and-error vocal learning. The involvement of dopaminergic neurons in this process and their known reward prediction error coding suggest that young songbirds maximize and an intrinsic a reward signal given by the similarity between their immature song and the memorized template.
In controlled tutoring experiments, we estimate the distribution of intrinsic reward as a function of acoustic pitch, a song feature that birds are known to copy from the template. We find that on the level of the song repertoire, intrinsic reward maximization bears resemblance to certain transport metrics used in natural language processing for estimating document similarity, where song syllables are analogous to words and syllable pitch is analogous to a word’s embedding.
Our work shows that based on normative reinforcement learning theories, we can estimate the intrinsic reward underlying complex skill learning. Songbirds can thereby provide an important link between vague notions of intrinsic reward in psychology and ideas on intrinsic reward in artificial agents playing computer games.
Professor Richard Hahnloser
Institute of Neuroinformatics,
University of Zurich and ETH Zurich
Zoom link: https://oist.zoom.us/j/98260915981?pwd=dWNFRVQrcUhCNWhrbGhtYWs0TEZPUT09
Meeting ID: 982 6091 5981