Pattern Separation in the Ventral Visual Stream
Abstract
Pattern separation is a neural computation thought to underlie our ability to form distinct memories of similar events. It involves transforming overlapping inputs into less overlapping outputs. In the ventral visual stream (VVS) there is considerable evidence for hierarchical transformation from feature-based visual representations to conjunctive whole-object representations, with the latter allowing for distinct coding even when objects have significant feature overlap. In the current study, we asked whether this transformation can be understood as pattern separation, and whether pattern separation can be observed even outside the context of classic recognition-memory tasks. To investigate pattern separation in the VVS, we combined fMRI in humans (N=23) with multivariate pattern analyses techniques and compared representations of visual objects in a mid-level visual region, Lateral Occipital (LO) region, with those in the region proposed to be at the top of the VVS object processing hierarchy, Perirhinal Cortex (PRC). During scanning we presented images of objects from multiple categories, with differing degrees of visual similarity among exemplars during performance of an N-Back task. Imaging results obtained using classification revealed patterns in LO could be distinguished successfully for all categories and at the lowest level of visual similarity within category exemplars. In contrast, patterns in PRC could be distinguished at all levels of similarity within a category, but no successful category differentiations were found. Because patterns at higher levels of visual similarity are overlapping in LO, but can be differentiated in PRC, these results provide evidence for pattern separation in the VVS. More broadly, this suggests that the engagement of pattern separation may not be restricted to the hippocampus during declarative-memory tasks.