Fig. 6.

Simulation of one-dimensional array of columns and voxels. Each column was assumed to respond to orientation input according to a Gaussian-tuning function peaking at its preferred orientation (SD, 45°; noise was added to the output). The preferred orientation shifted by a constant degree (a) or by a constant degree plus noise (b). In each trial, a single orientation was given as input, and the outputs of 100,000 columns (color band) were sampled by 100 voxels (gray boxes). The actual location of voxel sampling was randomly jittered on each trial (Gaussian distribution with an SD of a 1/4 voxel size) to take into account residual head motion. The number of stimulation trials was chosen to match the fMRI experiment. The sampled voxel data were analyzed using the same decoding procedure. As can be seen in the polar plots on the right, orientation can be readily decoded from the irregular array of columns (b), but not from the regular array (a). Similar results were obtained with a wide range of simulation parameters. Note that if voxel sampling is no longer jittered to mimic minor brain motion, orientation can be decoded even from the regular column array. This is because the high spatial frequency component can still persist after the sampling by large voxels. However, given that pulsatile brain motion and minor head motion cannot be fully eliminated or corrected with 3D alignment procedures, it seems unlikely that such high-frequency information contributes much to the orientation content in our fMRI data.