Abstract
Face pareidolia is seeing an illusory face in a non-face image. Recent studies have suggested that macaque monkeys also perceive such illusory faces in non-face images (Taubert et al, 2017) and that pareidolia images produce stronger fMRI activations in macaque face patches compared to control images (Taubert et al, 2020). However, it is unclear which image properties drive pareidolia responsiveness. Here, we test the hypothesis that pareidolia selectivity depends on the arrangement of eye- or mouth-like features in a face-like configuration. To address this question, we recorded neuronal responses in the medial and anterior lateral face patches in four monkeys. We presented 100 face images, 100 pareidolia images and 100 matched control objects (one per pareidolia image). For each image we also presented a “scrambled” version by randomly rearranging the four image quadrants, such that the individual features were preserved but the image was no longer in the original configuration. Overall, we found a higher response to pareidolia images compared to the matched controls and found that neurons that were more face selective were also more pareidolia selective. Both the neural preference for pareidolia and its relation to face selectivity could be captured by a deep neural network encoding model trained on a separate set of object images, suggesting that the face and pareidolia selectivity could be explained by common features. Finally, quadrant scrambling of the images reduced, but did not entirely abolish, either pareidolia or face selectivity. This suggests that the individual facial properties, as well as the spatial arrangement of features, contribute independently to pareidolia selectivity. In conclusion, we show that pareidolia and face selectivity in macaque IT responses are closely related and that they depend on both global feature arrangement and individual features.