A core aspect of social perception is thinking about people in terms of their social group memberships. Perceiving an individual in terms of their social category membership allows perceivers to apply pre-existing information and attitudes about that social category to the individual.

By providing information about otherwise novel individuals, social categorization reduces uncertainty about the social world and helps perceivers anticipate the characteristics of others and calibrate their own behavior. Although social categorization is necessary to make informed inferences about novel people, its reliance on cognitive shortcuts such as heuristics, stereotypes, and prejudices may lead to systematic inaccuracies in the processing of any given person.

In this line of work, we examine how stereotypes shape the predictions that deep neural networks make about people when social categories can be used to make inferences about people. For example, the very act of having structured group labels may bias early perceptual hidden layers to exaggerate features that define the labeled groups, and may “fill in” missing perceptual details (e.g., a gun instead of a tool).

By varying the nature of the learning environments, such as by limiting exposure to certain groups, or by establishing cultural myths about social groups (see project 1) we can understand how rational cognition can give rise to inaccurate and immoral beliefs in biased contexts. 

SocialAI Research Group