Cognitive Science Colloquium - Alexander Todorov / Modeling subjective visual evaluations
Cognitive Science Colloquium - Alexander Todorov / Modeling subjective visual evaluations
Thursday September 25, the Cognitive Science Colloquium welcomes Alexander Todorov from the University of Chicago, to share his work on modeling subjective judgments of visual stimuli, finding reliable statistical patterns despite the abstractness and variability of the judgments.
People effortlessly and spontaneously evaluate stimuli in their environment. There are two challenges understanding these evaluations. The first challenge is that reducing a complex evaluation (e.g., “beautiful”) to the physical description of the stimulus is far from trivial, because the space of hypotheses of what perceptual features drive evaluations is infinitely large. My research group has developed data-driven computational methods that allow us to find a consistent mapping from features to evaluations. Although the original methods were developed for face evaluation, modern machine learning methods extend the approach to modeling evaluations of any visual category. The second challenge is that evaluations are highly idiosyncratic. In fact, statistical modeling shows that stable idiosyncratic preferences account for most of the variance of complex evaluations (e.g., more than 80% in the case of “trustworthiness”). These findings suggest that the mapping from features to evaluations is highly heterogeneous across people. Yet almost all existing models of evaluations are models of aggregated judgments, assuming consistent mapping across people and, essentially, masking idiosyncratic differences. I describe one approach of building models of evaluations of individual participants. The models are meaningful and reveal the diversity of human preferences.