Title: Revealing hidden biases in face representation via deceptively simple tasks
Stefan Uddenberg, Principal Researcher, Booth School of Business, University of Chicago
Future Assistant Professor in the Department of Psychology at University of Illinois Urbana Champaign
Abstract: When we look at someone’s face, we can’t help but ‘read’ it. Within the blink of an eye, we can extract someone’s demographic characteristics (e.g., age) and transient psychological states (e.g., emotion). Not only that, but we form robust impressions of what we think someone is like as a person (e.g., how trustworthy they seem) — regardless of how (in)accurate these impressions may be. My work explores the roles of perception and memory in such feats in two complementary ways. First, I have discovered the existence of default face representations within our minds, and I am actively investigating their nature. These defaults are essentially unconscious and generic assumptions about facial features that bias or “pull” the representation of (and memory for) individual faces towards them. Second, I have applied deep learning methods to develop hyper-realistic generative models of human faces that vary along trait dimensions of psychological interest, such as perceived trustworthiness and dominance. Unlike previous computer models used in the field, these faces are nearly indistinguishable from actual photos; but unlike actual photos, they can be systematically and easily manipulated. I am currently using them to investigate a variety of research questions, such as how impressions of leadership differ with political affiliation. In this way, my work serves to bridge many different parts of our field and its history — drawing inspiration (and contributing to) both cognitive and social psychology, as well as bleeding edge revelations in the realm of machine learning. Collectively, these projects demonstrate (and leverage the fact) that our perceptual systems are geared toward extracting social properties from the stimuli in our environment.