Abstract
The neural representation of visual motion perception has been extensively studied in cognitive and visual neuroscience. Functional magnetic resonance imaging (fMRI) is often used in combination with multivariate pattern analysis to identify brain areas associated with motion perception (Kamitani&Tong, 2006). The assumption is that certain voxels are sensitive to motion direction, and the resulting activity pattern can be exploited to discriminate between alternative motion directions from new data. An alternative approach consists in specifying a forward model describing the mapping between changes in motion direction and the expected voxel activity with a basis set. The model estimates can be inverted and used to perform stimulus reconstruction (Inverted Encoding Models; Brower&Heeger, 2009). Encoding models typically seek the ideal response profile of motion-selective neuronal populations tuned to different directions. The choice of basis functions is often difficult, as cells tuned to motion direction can exhibit various of response profiles (Albright, 1984). Here we tested a novel nonparametric approach to motion direction reconstruction. The method is based on a cyclic version of Gaussian Process Regression (GPR – Rasmussen & Williams 2006) to obtain a continuous estimate of direction-dependent voxel responses. 24 participants performed a feature-continuous perceptual decision-making task during an fMRI experiment. In each trial, they viewed a 2s motion dot stimulus with different coherence (0%, 100% and a medium level) and direction (randomly varying from 0° to 360°), and indicated the perceived motion direction. For each subject, we estimated the trial-wise activity of individual voxels during the stimulus period. Using GPR, we obtained a continuous response profile for each individual voxel. The estimated voxels response profiles within a searchlight were then combined to obtain a trial-wise predicted direction. We found that the motion direction could be reconstructed from early visual cortex, and that the reconstruction was less precise with decreasing motion coherence.