FriSem

Eshed Margalit, PhD student in the Neurosciences Graduate Program, Stanford Medicine (working with Professors Dan Yamins and Kalanit Grill-Spector)

Title: Navigating Cortical Maps with Topographic Deep Neural Networks

Abstract: 1,414 days ago, I stood before the FriSem audience and presented preliminary results from an ambitious project to model the structure and function of primate visual cortex. In the weeks and months that followed, we would come to realize that most of the results presented that day were probably wrong. In this talk, I’ll describe our efforts over the past four years to construct strong, rigorously-tested models of visual cortex with topographic deep convolutional neural networks (DCNNs).

While DCNNs trained on object recognition tasks have been shown to impressively predict neural response properties, they have no spatial layout for features at a given retinotopic location and are thus unable to predict the rich topographic organization of visual cortex. In primates, this organization includes including pinwheel-like arrangements of orientation-tuned neurons in primary visual cortex (V1) and patches of category-selective neurons in higher visual cortex. Our approach, based on the simple principle that nearby neurons should be correlated during development, yields topographic DCNNs that simultaneously predict topographic properties and neuronal response properties in the ventral visual stream (but for real this time). I will discuss the biological phenomena this approach is able to reproduce, the surprisingly narrow set of constraints required to achieve those results, and the challenges that remain to be solved en route to models that reliably predict the structure and function of primate visual cortex.

You can find this information on the talk here. http://frisem.su.domains/february-4-2022-eshed-margalit/