Figure 2. Voxel-wise encoding model performance and voxel-binning procedure.

A) Histogram of voxel-wise encoding model prediction accuracy for all voxels in the functional volume acquired for Subject 1. For each voxel the model prediction accuracy is the correlation between the encoding model predictions and activity measured during the model-testing runs. The histogram has a mode near 0 and a heavy tail that is more easily appreciated on a log scale (inset). B) Overlay of model prediction accuracy onto a single axial slice (Subject 1). Voxels in which activity is poorly predicted (shown in blue) are scattered throughout white matter across the posterior-anterior extent of the scanned area. Voxels in which activity is accurately predicted (shown in pink) are confined to gray matter and thus track the convolutions of the cortical surface. C) To facilitate the image identification analyses voxels were rank-ordered by model prediction accuracy and then binned into populations each containing 1000 voxels. Populations are illustrated as circles surrounding schematized voxels (squares) whose color indicates model prediction accuracy. Low-rank populations (blue) contain poorly tuned voxels that had low prediction accuracy. High-rank populations (pink) contained well-tuned voxels that had high prediction accuracy. The lower (upper) bound on prediction accuracy for a population is determined by the voxel in the population with the lowest (highest) model prediction accuracy. D) The percent of early visual area (EVA; V1 and V2) voxels in each of the voxel populations used for image identification. The x-axis indicates the lower bound on model prediction accuracy for each population. The left y-axis indicates the percentage of the voxels in EVA across all populations with a lower-bound greater than or equal to the value on the x-axis. The right y-axis indicates the percentage of voxels in EVA in each population relative to the percentage of voxels in EVA in the functional volume (~2% for all subjects). The black curve shows how the percentage increases as the lower-bound on model accuracy increases. In the population with the highest lower bound > 20% of the voxels are in EVA. This is greater than 9 times the percent of voxels in EVA contained in the functional volume. The dashed line indicates the absolute (left y-axis) and relative (right y-axis) percentage of voxels in EVA that would be obtained if populations were constructed by randomly sampling voxels from the functional volume.