Reverse hierarchy theory began with an important insight from visual learning: training guided by generalities and involving easy discriminations was more efficient than training based on specific, detailed discriminations (Ahissar & Hochstein, Proceedings of the National Academy of Sciences, 90 (12), 5718–5722, 1993; Nature, 387 (6631), 401–406, 1997). This later developed into a more general theory concerning the flow of neural signals through the anatomical hierarchy (Hochstein & Ahissar, Neuron, 36 (5), 791–804, 2002). In brief, the contents of visual experience are assembled unconsciously on the way up; conscious experience begins from the top and works its way down. Here we apply this framework to social perception, where the conventional view suggests that social learning begins with detailed, observable behavior. Facial expressions, gestures, and gaze direction are registered and used to infer higher-level mental states, including goals, intent, and beliefs. In this review, we consider a reverse flow, with recursive updating. A prediction cycle begins with an assumption about another’s mental states, in a given social context, which is used to interpret the observable behavior. Mismatches between predictions and observables contribute to updated hypotheses in a recursive fashion. We show how this framework offers a solution to three paradoxes in social perception research: (1) preserved action imitation in autism, (2) definitional problems of empathy, and (3) coordinating self- and other-attributions in joint task performance. We claim that adopting this framework will improve our understanding of social perception through new studies with readily testable hypotheses and that it will inspire studies with greater ecological validity.
