2026 Theses Doctoral

Mapping the representational geometry of human face perception with generative models of face space

Bossio Botero, Veronica

With a single glance at a face we can make effortless inferences about a person's identity, age, and emotional state, often in a fraction of a second. Understanding the representations and computations that give rise to these abilities requires formal models of the perceptual space in which faces are represented that can generate testable predictions about human behavior.

In this thesis, I use two generative models of face space to probe two perceptual boundaries that are central to face processing: the boundary between faces that appear "human-like" and those that do not (Chapter 3), and the boundary between individual identities (Chapters 4 and 5). In the first case, I find that human judgments of face naturalness are highly sensitive to the direction in face space along which a face deviates from the average, not only to the magnitude of the deviation.

In the second case, I find remarkable idiosyncrasy in where individual observers draw the boundary between identities in face space. Building on this finding, I use closed-loop adaptive psychophysics to systematically identify face stimuli that maximally drive population-level disagreement about perceived identity.

Finally, I present a methodological contribution to the study of representational geometry beyond face perception (Chapter 6), clarifying the circumstances under which representational distances estimated from measurements of brain activity reflect the true representational geometry of the underlying neural population. Collectively, this thesis constitutes a systematic and quantitative investigation of the representations underlying perceptual judgments of face naturalness and identity.