From images to meaning: What have deep neural networks taught us about the ventral stream?

How does the ventral visual stream turn patterns of light into meaningful objects? In my lab, we combine experimental and computational approaches to characterize how object information is represented in human high-level visual cortex. We have shown that the representation is at once categorical and continuous: response patterns encode visual features of intermediate complexity (e.g., eye, round) that are diagnostic of ecologically relevant categories such as faces. Feedforward deep neural networks approximate aspects of this representational geometry, suggesting that relatively simple linear and nonlinear transformations of visual input can produce category-aligned structure. At the same time, our work reveals systematic divergences between artificial networks and human vision, particularly under more challenging and naturalistic conditions. I will argue that deep neural networks are most powerful not as end-point models of the ventral stream, but as controlled systems for probing the learning pressures and constraints that shape human visual representations.

Alternate locations:

Online: https://utoronto.zoom.us/j/85899245173