Speakers > Richard Hahnloser

Richard Hahnloser

ETH Zurich, Zurich, Switzerland

     Richard Hahnloser studied Physics and is Professor and Co-director of the Institute of Neuroinformatics, jointly affiliated with the University of Zurich and ETH Zurich. He studies vocal production and vocal learning in songbirds and is interested in sensory and motor systems and how they interact to solve a computational task. He is affiliated with the Neuroscience Center Zurich (ZNZ) and with the Swiss National Centre of Competence in Research (NCCR) Evolving Language.                    

                                 Reinforcement learning of birdsong

Reinforcement learning (RL) is a promising computational theory for understanding learned skilled behavior and its underlying neural mechanisms. Yet, most natural behaviors are too variable to be compatible with optimal action policies that strictly maximize reward. To reconcile these views, we postulate that optimality in reinforcement learning applies not to behavior itself but to a latent source of motor variability, from which an animal learns. We assume that this source is an ideal source of randomness, all other aspects of behavioral variability we attribute to non-ideal variability sources from which an animal cannot directly learn.

We test this hypothesis for brain organization in songbirds subjected to a vocal pitch reinforcement task. Our model agrees with a wealth of data and produces excellent fits to pitch data even when learning trends are non-monotonic due to circadian fluctuations. When we lesion the output area of the known pitch-learning circuit, we find an excellent linear relationship between lesioned volume and reduction of behavioral variability attributed to the latent learner. Estimation of learner variance is possible even from spontaneous unreinforced behavior, which provides a convenient access to a localized brain function.