Events

On Adversarially Robust Models of Vision, "Brain-Alignment'', and their use for Understanding Visual Art

Speaker: Arturo Deza

Location: 6 Washington Place, Room 159
Videoconference link: https://nyu.zoom.us/my/ajay.subramanian

Date: Monday, March 27, 2023

One of the key differences between modern machine vision systems and human perception is their disalignment when interpreting complex imagery. Over the last decade, engineers & psychologists have been focusing on how to solve such alignment problems when machine vision systems are exposed to their adversarial imagery (different from that of humans), in addition to out-of-distribution stimuli. In this talk, I will present two recent works that explore: 1) a natural solution in human perception that may lead to adversarial robustness via testing of complex stimuli in the visual periphery; 2) a strange emergence of "brain-aligned" representations in a particular type of Transformer architecture that strongly correlates with primate activations in area V4. I will then finish the talk with real world applications of "brain-aligned" machine vision systems and how they are currently being used to power reverse image search engines of Visual Art and other complex imagery.