Abstract
We present a novel methodology that exploits work in computer vision, for identifying the neural correlates of face perception. It involves creating a sequence of non-face natural images that, on the one hand, constitute a smoothly varying continuum in terms of image-level similarity to faces, and on the other hand yield categorical face/non-face responses from observers. We then determine which brain regions exhibit graded activity changes across the sequence and which are categorical.
The key challenge in operationalizing this approach lies in compiling the collection of natural images that would constitute a continuum in a face-similarity space. We have addressed this challenge by using a state of the art machine-based face detection system. Given any image, the system spots faces therein and, on rare occasions, generates false positives (FAs). We have collected several hundred such FAs and rank-ordered them in terms of their similarity to faces (using computational norms and also ratings from human observers). The final collection comprises 300 images, ranging smoothly from being very distinct from faces to genuine faces. Similarity ratings change almost linearly across this set, while the face/non-face labels change categorically.
Using fMRI, we investigated which brain regions correlate with categorical face/non-face judgments. As a preliminary test of the approach, an ROI analysis showed that fusiform face area exhibited perceptually congruent categorical responses. In contrast to high level activity induced by faces, all non-face images, irrespective of their similarity to faces led to low activations. These results illustrate of how the approach allows us to assess whether a given region's activity is involved in categorical judgments rather than just being correlated with low-level image structure. We shall describe the use of this approach with whole-brain responses to identify additional regions involved in face processing.
Research supported by The Simons Foundation