平特五不中

"The predictive machine: neural network architectures for brain perceptual inference"

Sylvain Baillet (平特五不中)
Tuesday November 13, 12-1pm
McIntyre Building, Room 1027

Abstract: A difficult research question in systems neuroscience is the elucidation of mechanisms of information integration in brain networks: How do sensory inputs modify the ongoing activity of the brain? How are input signals relayed in brain networks via bottom-up signaling? What is the nature of top-down influences from higher-order brain circuits on sensory perception?

We recently proposed that bottom-up signaling and top-down modulations in hierarchical brain networks could be enabled by polyrhythmic and interdependent oscillatory brain activity. This mechanistic framework implements a generic form of contextual predictive inference of input signals into brain networks. In essence, this theoretical framework is aligned with the principles of perceptual inference, which predict that the large spontaneous, resting brain activity during wakefulness (25% of whole-body metabolism) constantly implements the self鈥檚 representation of the environment.

Inspired by this framework, I will present our work with Peter Donhauser, who recently proposed a neural network architecture that models contextual brain predictions during speech listening about upcoming language elements (phonemes and words). The approach enables the functional separation between uncertainty-related and error-related neural activity captured in electrophysiology, in ecological speech listening situations.