Abstract
To encode and make inferences about the world, the brain represents patterns -- visual, auditory, cognitive -- using populations of neurons with diverse and complex forms of selectivity. A coarse-scale method like functional magnetic resonance imaging (fMRI) would, at first glance, appear poorly suited to studying these representations. Consider the case of orientation representation in primary visual cortex (V1). A single fMRI voxel pools responses from many orientation-tuned neurons. Because orientation tuning varies at a fine, columnar spatial-scale, tuning should cancel at the level of fMRI voxels. Surprisingly, results from multivariate decoding analyses imply that voxels in human V1 are weakly but reliably orientation selective. It is widely believed that these small biases arise because of random spatial irregularities in the underlying columnar architecture, and this interpretation, while untested, has been extended to the study of cognitive functions throughout the brain. I will describe a set of experiments that test this hypothesis by characterizing both orientation selectivity and motion direction selectivity using fMRI in human visual cortex. We discovered coarse-scale biases for both orientation and motion direction. The existence of these coarse-scale biases are both necessary and sufficient for multivariate decoding, demonstrating that random spatial irregularities do not contribute to decoding. Our results imply a parsimonious, but sobering, explanation for why fMRI decoding works, and help guide the interpretation of the rapidly growing number of studies based on this technique.