Conference on Cognitive Computational Neuroscience

The ability to recognize emotions and intent in facial expressions varies significantly between neurotypical (NT) individuals and those with autism spectrum disorder (ASD). Traditional inferential models often utilize high-level categorical descriptors of stimuli, neglecting the variance introduced by image-level sensory representations. This study investigates whether accounting for image-level differences can improve the development of diagnostic image-sets for emotion recognition tasks. We employ image-computable artificial neural network (ANN) models of primate vision, fine-tuning them on an existing dataset to predict the behavior of NT and ASD adults. Using these ANNs, we select a new set of images that are predicted to yield the largest differences in performance between NT and ASD subjects. Subsequently, we conduct facial emotion discrimination tasks and find that the ANN-selected images produce significantly larger behavioral gaps between the groups compared to a random selection of images. Notably, the diagnostic efficiency of the selected images can be predicted by the ANNs' ability to predict NT subject behavior. Our findings suggest that ANN models of vision could offer valuable clinical translation benefits for autism research, opening up new avenues of exploration.