Abstract
We used fMRI in conjunction with pattern classification to characterize the neural mechanisms responsible for the attentional selection of complex objects. Stimuli consisted of single faces, single houses, and blended images in which the two images were spatially overlapping. Observers performed a same/different discrimination task involving pairs of sequential stimuli. In the blend condition, observers performed the discrimination task for one object type, which required attending selectively to either faces or houses while ignoring the other object type. Functional activity patterns from individual visual areas (V1-V4, fusiform face area, and parahippocampal place area) were used to train a linear classifier to predict the object category that was seen or attended on independent fMRI test blocks. Activity patterns in both high-level object areas and low-level retinotopic areas could accurately distinguish between single faces and houses. More importantly, activity in these areas could reliably distinguish the target of attentional selection in the blended stimulus condition, across all visual areas tested. When classifiers trained on single object conditions were used to predict the attended object category in the blended condition, decoding performance was high across all areas indicating wide-ranging effects of object-based attention. Furthermore, block-by-block analysis of the strength of the object-specific attentional bias revealed a strong correlation between bias signals in object-selective areas and low-level areas. Finally, the object-specific activity patterns found for attended upright objects could effectively generalize to cases when the same object classes were viewed upside down. Results indicate that an object-selection mechanism involves a multi-level bias modulating diverse feature responses throughout the visual pathway. Selecting one of two overlapping objects involves widespread biasing of cortical activity, resulting in activity patterns that resemble the target object viewed in isolation. Such attentional filtering may be essential for flexible efficient processing in crowded and complex real-world visual settings.
NIH R01-EY017082 and NSF BCS-0642633.