Abstract
Multivariate pattern analyses are widely used in neuroimaging studies. But it is unclear what these analyses reveal about neural coding. Such approaches often obfuscate the link between neural measurements and the stimulus, greatly limiting their utility. Linear classification, a highly-popular approach, exploits the Euclidean distance among pattern vectors. Our recent theoretical work suggests that information regarding tuning properties is best captured by angular distances, not Euclidean distances. The goal of this experiment was to relate specific aspects of multivariate measurements to parametrically defined dimensions of visual stimuli. We used primary visual cortex as a model system to test the following predictions: (i) Euclidean distances conflate information regarding response amplitude and selectivity, (ii) angular distances best reflect information about selectivity, and (iii) data demeaning—a common preprocessing procedure—invalidates inferences regarding selectivity because it changes angular relationships among pattern vectors. fMRI BOLD activity was measured from 5 participants (7T, 32-channel coil, 1.2x1.2x1.2 mm resolution). Stimuli consisted of Cartesian sinusoidal gratings multiplied by flower-shaped apertures. The orientation of the “petals” was controlled by the phase of the radial frequency (RF) used to generate the aperture. In a fully crossed experimental design, the compound stimulus varied along three dimensions: grating orientation, contrast, and RF-phase, resulting in a total of 96 unique conditions. We found that the Euclidean metric was sensitive to stimulus contrast and relatively insensitive to grating orientation and aperture phase. Angular distances proved sensitive to grating orientation and aperture phase, and robust to changes in stimulus contrast. As predicted, data demeaning led to mischaracterizations of neural tuning. Finally, we used a biologically-inspired image-computable model to provide a theoretic account of these observations. Our results have fundamental implications for the interpretation of MVPA, and inform the development of data analysis strategies that are more readily interpretable in terms of the underlying neurophysiology.