Abstract
Neurons in early visual cortex receive highly selective feedforward input that is amplified or disamplified by contextual feedback and lateral connections (Phillips, 2015; Gilbert & Li, 2013). Many cortical areas feed back to early visual cortex, yet measuring feedback channels presents a central challenge to fully understanding neural computations (Muckli & Petro, 2013). To isolate feedback we blocked feedforward input to subsections of retinotopic visual cortex with a uniform visual occluder covering one quarter of the visual field (Smith and Muckli, 2010) while participants viewed 24 real-world scenes. Scenes spanned six categories and two spatial depths, allowing us to investigate whether feedback contains information about these features. By using pattern classification techniques we found that response patterns in occluded subsections of V1 and V2 contain individual scene, category and depth information, and that category information was generalizable across scenes while depth information was not (Figure 1). Feedback to early visual cortex is therefore specific to individual scenes while concurrently conveying some high-level structure. Still, occluded V1 and V2 responses differed from each other, suggesting that feedback to these two areas has unique information content (Figures 2 & 3). Finally, V1 and V2 represented scenes differently than three popular biologically-inspired computational models (Weibull, Gist, and H-MAX; Figure 3). Together, these results suggest that in order to understand neural computations of early visual cortex, we must understand the contribution of cortical feedback.
Meeting abstract presented at VSS 2016