Abstract
Human observers easily prioritize processing of task-relevant stimuli in cluttered visual scenes. This study used fMRI to characterize how effectively regions across human visual cortex filter a target's spatial location or content from task-irrelevant distractors. Nine subjects performed a task that involved following spatial cues to attend content (faces or houses, 9 degrees visual angle in diameter) in one of two parallel stimulus streams (5.1 degrees above or below fixation) in 16.2s blocks. We used cross-classifiers to obtain separate readouts of content-tolerant spatial priority representations and spatially-tolerant content representations. To isolate the efficacy of top-down attentional filtering from other factors that influence absolute classification accuracy we defined a filtering efficacy metric as the ratio of classification when the distractor stream was present compared to absent. Thus, filtering efficacy is 1 when classification performance is unaffected by the distractor and 0 if the distractor reduces classification to chance. Both dorsal and ventral visual stream regions were found to feature highly effective content filtering (efficacy>0.75 in V3a, TOS, ips0-2, hV4, OFA, PPA and FFA). Spatial filtering also was similar along a dorsal-ventral axis, but was maximally effective in mid-level visual cortex (efficacy>0.66 in V3, V3a and hV4). By contrast, the classic where-what division was evident in the profile of classification accuracy for bottom-up visual responses in the no distractor context, with most accurate spatial classification in dorsal regions and content classification in ventral regions. Our results suggest that spatial priority- and content-based filtering are distinct attentional mechanisms which are dissociated based on their efficacy profiles along a posterior-anterior axis. Thus, top-down filtering acts diffusely across both streams rather than following the organization of bottom-up response preferences into dorsal 'where' and ventral 'what' streams.
Meeting abstract presented at VSS 2014