Recent claims about the superiority of deep-nets to model human vision are based on the reproduction of either (a) neural recordings from different brain layers or (b) high-level behaviors included in Brain-Score. However, high correlations in such tasks do not guarantee the functional similarity of the underlying mechanisms in models and humans. In particular, appart from exceptions (like your work ;-), not many people is looking at the bottleneck of artificial systems as characterized by low-level visual psychophysics. In this talk we present stimuli for 20 different experiments that highlight basic color/texture/motion perception facts, and how the trends of the artificial responses can be used to assess the similarities with human vision.
Recommended reading:
* Li, Gomez-Villa, Bertalmío & Malo
Contrast sensitivity functions in autoencoders
Journal of Vision (May 2022), Vol.22, 8. doi:https://doi.org/10.1167/jov.22.6.8
* Malo, Vila-Tomas, Hernandez-Camara, Li, Laparra
A Turing Test for Artificial Nets devoted to model Human Vision (june 2022)
http://isp.uv.es/docs/talk_AI_Bristol_Malo_et_al_2022.pdf
* Malo & Bowers
The Low-Level vision MindSet.
Talk at the Seminars on Generalisation in Mind and Machine. (sept. 2024)
http://isp.uv.es/docs/Low_Level_MindSet_3.pptx
Zoom link: https://cam-ac-uk.zoom.us/j/81459588387?pwd=Uyq625q1QSKJ15ZvvSnzAbV44NTf1w.1
