Abstract
The organization of representations in the ventral visual stream (VVS) is thought to be hierarchical, such that posterior VVS represents simple object features, whereas anterior VVS supports increasingly complex conjunctive representations of multiple features. Despite considerable empirical support for this representational hierarchy for processing novel objects, it is unclear what changes occur to distributed object representations with extended learning. The perceptual expertise literature shows that discrimination between complex objects becomes faster with experience; this is a hallmark of unitization theory, whereby multiple features can be unitized and accessed as rapidly as a single feature. Keeping the organizing principle of the representational hierarchy in mind, this simple idea makes a powerful and unique prediction: unitization through perceptual training should modify conjunctive representations, but not simply as response tuning of existing representations. Rather, conjunctive representations would be redistributed to posterior VVS, whose architecture is specialized for processing single features. To test this hypothesis, we used fMRI to scan participants before and after visual training with novel objects comprising 1-3 features that were organized into distinct feature conjunctions. First, we used neural pattern similarity to replicate earlier findings that complex feature conjunctions were associated with conjunctive coding in anterior VVS. Critically, we also demonstrated that for well-learned objects, the strength of conjunctive coding increased post-training within posterior VVS. Furthermore, multidimensional scaling revealed increased pattern separation of the representation for individual objects following training. Finally, we showed that functional connectivity between anterior and posterior VVS increased for unfamiliar objects, consistent with early involvement in unitizing feature conjunctions in response to novelty. While there is strong behavioral support for unitization theory, a compelling neural mechanism had been lacking to date. Here, we leveraged recent advances in VVS subregional function to link established behavioral observations with representational transformations in the human brain.
Meeting abstract presented at VSS 2017