A bottom-up approach to finding individual differences in mental representation

Poster Session C - Sunday, April 14, 2024, 5:00 – 7:00 pm EDT, Sheraton Hall ABC

Y. Ivette Colón¹ (ycolon@wisc.edu), Claire Peplinski¹, Timothy Rogers¹; ¹University of Wisconsin - Madison

Individual differences in mental representation are often studied by first defining different groups, then assessing whether their representations differ in important ways. We consider a more data-driven approach to discovering representational subgroups without grouping participants beforehand. We employ a triadic judgment task to measure the semantic similarity of a set of DallE-generated faces, places, and objects that systematically vary in five binary attributes per domain (N = 118). For each domain, we computed 5-dimensional item embeddings for each individual, and for the group as a whole. The group embedding reliably encoded all five binary attributes in each domain (e.g. for face images, embeddings reliably differentiated race, perceived gender, age, setting, and time-of-day). To find representational subgroups, we first computed, for all pairs of participants, a representational similarity score by taking the Procrustes similarity of the corresponding individual embeddings. We then clustered individuals based on these similarities. For each distinct cluster of respondents we calculated cluster-level embeddings and assessed what properties these encode. We found subgroups that differ significantly in their conceptual representations of faces, mainly varying in the weight given to the different latent attributes (race, gender, age, context, and time). Fewer distinct groups were observed for representations of places and objects, suggesting that we vary more consistently in how we view the social world over the places we visit and things we use.

Topic Area: LONG-TERM MEMORY: Semantic