Deep learning models for the perception of human social interactions

Abstract

Social interaction perception is an important part of humans' visual experience. How- ever, little is known about the way the human brain processes visual input in order to understand social interactions. In comparison, other vision problems, such as object recognition tasks, have been studied extensively and seen success by comparing state of the art computer vision models to neuroimaging data. In this thesis, I employ a similar method in order to study social interaction perception with deep learning models and magnetoencephalography (MEG) data. Specically, I implement dierent deep learning computer vision models and test their performance on a social inter- action detection task as well as their match to neural data from the same task. I nd that detecting social interactions most likely requires extensive cortical process- ing and/or recurrent computations. In addition, I nd that experience with action recognition does not improve social interaction detection.