Speaker: Prof. Luca Mazzucato (Institute of Neuroscience, University of Oregon, USA)

When: 11 December 2024, h. 5:30 pm

Where: 1AD100, Torre Archimede

Abstract: Recent years have brought tremendous progress in the fields of artificial intelligence and
neuroscience, however, a comprehensive framework to explain the computational principles of
complex cognitive function and flexible behavior is still lacking. Addressing the complexity within
large-scale, multi-area neuronal recordings with simultaneous monitoring of behavioral variables
demands novel and deeper connections between the different axes of theoretical neuroscience. In
this talk, I will review current approaches and highlight challenges in the upcoming NeuroAI field
with particular focus on the following questions: i) What kind of dynamical systems can explain the
complex spatiotemporal features of naturalistic animal behavior? ii) What new physical ingredients
should recurrent neural networks incorporate to explain the richness observed in large-scale
neural activity during behavior? iii) What is the geometry of neural representations in animal brains
and how is it related to the geometry of deep learning? Finally, I will show some recent work in my
lab showing how the optimal cognitive function known as the Yerkes-Dodson law of psychology
relies on a phase transition occurring in sensory cortical circuits.