Search Papers | Poster Sessions | All Posters
Poster A122 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Designing salient, naturalistic “super-stimuli” with deep generative models
Shankhadeep Mukherjee1, Achin Parashar1, Chandra Murthy1, Devarajan Sridharan1 (); 1Indian Institute of Science, Bangalore, India
Attention can be deployed voluntarily, "top-down'', based on task goals, or captured automatically, "bottom-up'', by salient stimuli. Most previous studies have controlled stimulus salience by altering low-level target features, for example, by increasing luminance, or inducing popouts, or by creating motion dynamics. Can we design static, naturalistic images that capture bottom-up attention robustly? We address this question in three stages. First, we hypothesize that such tailor-made "super-stimuli'' would evoke strong responses in visual cortical areas. We advance a deep generative framework with a heuristic optimization algorithm (XDream) to design category-specific, naturalistic images that can combinatorially activate and suppress multiple regions across human visual cortex. Second, with human functional MRI recordings, we show that such super-stimuli differentially activate visual cortical regions targeted by the optimization algorithm, thereby validating the approach. Third, we show that, in a working memory task, super-stimuli optimized for specific regions are more accurately recalled than control stimuli, thereby demonstrating their behavioral salience. Our behaviorally validated super-stimuli open up new avenues of research for investigating neural mechanisms of exogenous attention control with salient, naturalistic images.
Keywords: super-stimuli deep generative networks fMRI attention