涩里番

Reverse-engineering cognitive and neural representations with multilevel computational theories

Ilker Yildirim, Yale University
Tuesday April 1, 12-1pm
Zoom Link:听
In Person: 550 Sherbrooke, Room 189

Abstract:听There lies a great distance between the raw inputs sensed at our retinas to what we experience as the contents of our percepts and thoughts --- objects with 3D shapes and physical properties, predictions of how these objects will move, and our plans toward them. What algorithms underlie these 鈥渙bject cognition鈥� abilities in the brain, including the formats of neural object representations, how these representations are inferred across the sensory cortex, and how they are used to predict what will happen next? Across two projects, I鈥檒l provide evidence that the brain implements object cognition by building and manipulating 鈥渨orld models鈥� 鈥� structure-preserving, behaviorally efficacious representations of the physical world. I鈥檒l do so by presenting 鈥渕ultilevel computational theories鈥�, which are computational models that natively interoperate in both the cognitive level of structure-preserving object representations and the neural level of distributed, multi-area codes and attractor dynamics. Using these models, the first project reveals a multi-area algorithm of the macaque inferotemporal cortex as reversing a graphics-based generative model of how 3D scenes form and project to images. The second project reveals neural mechanisms of mental simulation in macaque prefrontal cortex playing a variant of the game pong, as variable attractor dynamics implementing a world model of this game. Together, these studies start to reveal how the rich and generalizable knowledge of the world of objects is encoded in the brain鈥檚 biological neural networks.