Learning Dynamics in artificial and biological neural networks

Directors: Jean-Rémi King and Yves Boubenec

Jury: Maria Chait, Christophe Pallier, Micha Heilbron, Sharon Peperkamp and Emmanuel Dupoux

Résumé:  

Humans, in contrast to numerous species, are very sensitive to complex structures, for example, the repetition of a musical motif. To process these structures, they rely on the hierarchical organization of their sensory system. For auditory stimulus, the system first processes small chunks of audio into a set of features, which are then integrated. This integration is eased by the use of structures, which the brain automatically detects. Remarkably, our best models for how these detections unfold are artificial neural networks. These networks can predict the neural responses during sensory processing and can be used to decode from the brain a heard stimuli, like a sentence. As far as they can help us model how sensory processing unfolds, it remains unclear if they also model the emergence of these processes: its learning dynamic over several months. Therefore we studied the learning dynamic of  self-supervised models while comparing them with data recorded from Humans and Ferrets.