We will also stream this seminar via Zoom.
For the link, please contact us: pennmindcore@sas.upenn.edu
Tyler Bonnen
Presidential Postdoctoral Fellow
Electrical Engineering and Computer Science
UC Berkeley
Modeling neural function within a deep learning framework requires a mechanistic account of animal behavior
Over the last decade, optimization frameworks from deep learning have been increasingly adopted within the neurosciences. Still, it has proven challenging to apply these ‘stimulus-computable’ methods to a broad set of behaviors and neural structures. In this talk, I’ll suggest that mechanistic accounts of animal behavior play a critical – often underappreciated – role in modeling neural function within an optimization framework. As a case study, I’ll speak about my work characterizing perceptual behaviors that depend on the medial temporal lobe. By situating lesion and electrophysiological data within a deep learning framework, I was able to resolve decades of apparent inconsistencies in the experimental literature. These results implicate the medial temporal lobe in perception, providing unambiguous architectural constraints for models of human vision. It was only through a series of behavioral experiments, however, that it became clear there are algorithmic constraints that are just as important; remarkably, these are the processes that enable humans to outperform state of the art computer vision models. I’ll conclude by sketching out some of my ongoing work modeling these data within a biologically plausible optimization framework, alongside some general lessons that apply to modeling a richer space of neural structures and behaviors.
A pizza lunch will be served. Please bring your own beverage.