The Center for Philosophy of Science at the University of Pittsburgh invites you to join us for our Lunch Time Talk. Attend in person at 1117 Cathedral of Learning or visit our live stream on YouTube at https://www.youtube.com/channel/UCrRp47ZMXD7NXO3a9Gyh2sg.

LTT:  Aliya Rumana

Tuesday, March 11  @ 12:00 pm - 1:30 pm EST

Title: A deflationary account of DCNN-based models in visual neuroscience

Abstract: Deep convolutional neural networks (DCNNs) have achieved extraordinary accuracy at predicting electrophysiological data in the ventral visual stream (VVS). What explains these predictive successes? According to Cao & Yamins (2024), these models are so predictively successful because (a) they near-optimally perform the same tasks as the VVS (e.g., image classification) and (b) they perform these tasks in the same kind of way due to their shared mechanistic structure. For the latter reason, DCNNs are often touted as mechanistic models of the VVS. In this presentation, I’ll argue that a weaker version of the first reason is sufficient: these models are so predictively successful just because they near-optimally perform proper parts (approximately half) of the tasks that the VVS performs—not because they share any mechanistic structure. Any structural similarities between DCNNs and the VVS is incidental to their predictive success, so I conclude that DCNNs do not provide plausible mechanistic models of the VVS.

Can’t make it in-person? This talk will be available online through the following:

Zoom – https://pitt.zoom.us/j/96457118345

YouTube at https://www.youtube.com/channel/UCrRp47ZMXD7NXO3a9Gyh2sg.