To monitor the environment for social threat humans must build affective evaluations of others. These evaluations are malleable and to a high degree shaped by responses engendered by specific social encounters. The precise neuronal mechanism by which these evaluations are constructed is poorly understood. We tested a hypothesis that conjoint activity in amygdala and fusiform gyrus would correlate with acquisition of social stimulus value. We tested this using a reinforcement learning algorithm,
Q-learning, that assigned values to faces as a function of a history of pairing, or not pairing, with aversive shocks. Behaviourally, we observed a correlation between conditioning induced changes in skin conductance response (SCR) and subjective ratings for likeability of faces. Activity in both amygdala and fusiform gyrus (FG) correlated with the output of the reinforcement learning
algorithm parameterized by these ratings. In amygdala, this effect was greater for averted than direct gaze faces. Furthermore, learning-related activity change in these regions correlated with SCR and subjective ratings. We conclude that amygdala and fusiform encode affective value in a manner that closely approximates a standard computational solution to learning.