Fig. 1: Increasing pattern separability through top-down predictions.

A Similar faces, like those of lookalikes and twins are hard to distinguish because they lead to poorly separable neural activity patterns. Contextual predictions, derived, e.g., from previously learnt associations, can aid distinguishing similar stimuli. We hypothesize that predictive context facilitates recognition by increasing separability between neural activity patterns. E.g., a teacher who has trouble distinguishing identical twins in his class may utilize learnt contextual associations like friends of one of the twins to help distinguish the similarly looking faces. B In the face processing system, the lowest area ML has view-specific shape representations, the intermediate area AL has mirror-symmetric representations, and the top of the hierarchy, AM, has view-invariant appearance representations. All stages are directly and reciprocally connected. Higher face areas contain neural representations with minimal overlap (high separability) and view-invariance. We hypothesize that incorporating high-level predictive information (via feedback pathways) increases invariance and separability in lower areas, C To create contextual predictions, we exposed monkeys to face pairs in a statistical learning paradigm. Once learned, the first face in a pair (the “predictor”) predicts the second face in a pair (the “successor”). In the test phase, we presented predictable face pairs (60% of all trials) along with their violations during fMRI. Artwork depicting faces of kids in panel A by M. Shitik and is reproduced with permission by the artist. The image of the teacher is reproduced with permission from Alamy. The face images in panels B and C were created using FaceGen Modeller.