Richard Hahnloser: Can we formalize how animals explore the world? Unexpected insights from a binary framework in songbirds.

University of Zurich and ETH Zurich

Our lab is interested in identifying principles of natural intelligence using vocal behaviors in songbirds as a model system. One of the most compelling computational approaches to intelligence is the reinforcement learning framework. Reinforcement learning has been widely studied in birds using brief auditory or visual stimuli applied during singing. When such stimuli are applied contingently on the pitch of one of the bird’s song syllables, for example during low-pitched renditions of a syllable, birds tend to avoid the stimuli by increasing the pitch of just that syllable.

I will present our ongoing work of identifying the simplest descriptive model that can fully account of birds’ complex pitching strategies. Current reinforcement learning models are unable to account for the observation that birds are repelled by a stimulus in some situation and attracted by it in another. In particular, normally, birds tend to avoid pitch-contingent light off in their cage. However, our work shows that when deafened, birds change their song to increase the rate of light-off events. To account for this apparent value inversion of light off induced by deafening, we suggest an intrinsic motivation of birds to increase entropy. According to this principle, birds tend to choose actions that let them experience a maximally rich sensory world. We argue that our simple binary framework, deaf birds subjected to just two lighting conditions, is an ideal experimental setting for testing theories of motor exploration, which are of high interest in artificial intelligence where they are known as policies, but for which little is known in higher animals.

 

Guests are welcome!

 

Organized by

Michael Brecht / Margret Franke