Abstract
In this fMRI study we use multi-voxel pattern analysis (MVPA) to investigate the representation of emotional expressions in the superior temporal sulcus (STS). Subjects were shown a series of videos of morphed emotional expressions ranging on a 10-point continuum between pure fear and pure anger. The task required subjects to categorize the videos as one of these two emotions. For each morph level, we used sparse multinomial logistic regression (SMLR) to measure the similarity between the neural response during the perception of the morphed emotion and during the perception of a pure emotion. This can be accomplished by first training the SMLR classifier on the pure emotions. Then, when testing the classifier on each intermediate morph, the classifier produces a logit, which is a measure of how strongly the classifier believes the morph is anger as opposed to fear. We expected to see one of three outcomes: First, the relationship between morph level and logit could be linear. That is, as the morph level moves away from fear, the neural response moves linearly away from the representation of fear and towards the representation of anger. Second, the relation could be sigmoidal. Under this outcome, the difference in neural representation between two morphs within the same category is smaller than the difference between two equally spaced morphs that belong to different categories. Finally, as predicted by recent work on the STS (Said et al. 2008), the relationship could be nonmonotonic. Under this outcome, neural representations of pure emotions are closer to the neural representations of same-category morphs near the category boundary than same-category morphs further from the category boundary. Our results support this last hypothesis. We conclude that the STS can engage in sharp contrast enhancement at category boundaries, in a process that may help prevent confusion between competing categories.