Investigating the influence of priors in human fear generalization using a Bayesian model
Kampermann L, Büchel C & Onat S
When organisms form associations through learning, responses often generalize to stimuli that bare resemblance to the initially reinforced stimulus (CS+). During generalization, shifts in maximal responses away from the actual reinforced stimulus are commonly observed when tested with multiple stimuli along a similarity continuum, a phenomenon known as peak shift. We have developed a Bayesian framework that can account for peak shifts observed in behavioral ratings by postulating a group-level prior that is common to all participants over the range of stimuli forming the similarity continuum. The interaction of the prior distribution with a generalization component centered on the CS+ can act as a “magnet”, resulting in shifts of maximal responses depending on which stimuli was used as CS+. This formalism allows us to reverse-engineer subjects’ latent prior distributions by evaluating to what extent a specific hypothesis on priors can explain the observed shifts.
We employed a fear conditioning procedure using 8 faces organized along a circular similarity continuum varying in gender and identity dimensions. Two opposite faces were randomly selected as the CS+ and CS– for each participant (n = 141). Following aversive conditioning, we obtained fear generalization gradients by asking participants for explicit UCS expectancy ratings, thus obtaining a two-sided generalization gradient ranging from CS+ to CS–. We compared the performance of the Bayesian model to fits carried out on a single-subject basis using two versions of Gaussian functions and subsequently tested different hypotheses on priors for the Bayesian model.
As expected a flexible Gaussian model with 2 parameters (width and location parameters) fitted the data significantly better than a simple Gaussian model (only width parameter) centered on the CS+ face (r = .91 vs. r = .74) underlining the presence of peak shifts in behavioral ratings. On the other hand, the Bayesian model performed significantly better than the simple Gaussian model, despite being based on only two more free parameters (flexible Gaussian: 2n parameters; simple Gaussian: n parameters; Bayesian model: n+2 parameters). Furthermore the Bayesian model explained behavioral gradients best, when a bimodal prior distribution peaking at both gender prototypes was used (r = .84). Testing different gender categories individually, a unimodal prior centered on the male category explained as much variance as one centered on female category. Overall, the predominance of a bimodal prior indicates that peak shifts can result from “magnet” effects of categorical face representatives instead of adversity attributions to a specific gender.