Academia Sinica Logo

What can brains teach machines, and what can machines teach us about brains? This talk presents a framework for NeuroAI built on the bidirectional exchange. We first explore what AI can tell us about how the brain is organized for vision; in particular, what loss functions and architectural constraints best explain the selectivity of visual neurons for objects and motion. We then reverse the arrow: given that neuroscience moves slowly while AI capabilities accelerate, how should we strategically deploy brain-inspired ideas? The answer lies in targeting persistently gnarly problems, like adversarial robustness and AI safety, where biological solutions offer durable advantages. Finally, we close the loop with a vision for the automated neuroscientist: a system pairing an outer-loop reasoning agent with an inner-loop brain model to accelerate discovery.

Bio:

Patrick Mineault is the NeuroAI lead at the Amaranth Foundation, which funds ambitious research in neuroscience. He completed his PhD on the computational neuroscience of vision at McGill University. He was a data scientist and software engineer at Google, a brain-computer interface engineer at Meta, a staff scientist at Mila, and was one of the cofounders of Neuromatch. He writes about NeuroAI on neuroai.science