Abstract
Multi-voxel pattern analysis (MVPA) is typically treated as a black box. Studies generally focus on whether stimuli can be decoded from a brain representation, while the source of decodable information is hidden within complex activation patterns. One notable exception is research aiming to understand how orientation information, represented in 1mm hypercolumns, can be decoded from visual cortex using fMRI, which measures brain activity at a coarser scale (~3mm). Some have argued that MVPA confers hyperacuity to fMRI by recovering fine-grained information derived from imperfect sampling of orientation columns within voxels. Others have argued that a coarse-scale radial bias is the source of information for orientation decoding. Recently, in modelling the cortical response to grating patterns we identified a third potential source of information, a non-uniform spatial response produced by the edges of the stimulus (Carlson, 2014). In the present study, we examined the contribution of edge-related activity to orientation decoding with fMRI. Subjects viewed annular square-wave grating stimuli of six orientations in a blocked design. Orientation was reliably decoded using a linear SVM. Localizers and retinotopic mapping were used to identify voxels corresponding to the inner edge, middle, and outer edge of the stimulus in retinotopic cortex. Analyzing these regions separately, we found that decoding performance was highest for the area of the cortex representing the inner edge of the annulus. We then used the method of Haufe et al. (2014) to transform the voxel weights from the classifier into interpretable activation space to examine the location of the voxels most significantly driving classifier performance. The majority of the highest-weighted voxels overlapped with the region of retinotopic cortex corresponding to the inner edge of the annulus. Our findings suggest that stimulus edges contribute to orientation decoding, and underscore the importance of considering the source of decodable information in MVPA studies.
Meeting abstract presented at VSS 2015