In recent fMRI studies, we have shown that activity patterns in early human visual areas contain robust orientation information that allows for accurate decoding of which of eight orientations is seen on a trial-by-trial basis (Kamitani & Tong, VSS/SFN 2004, Nat Neurosci, 2005). Our findings suggest that random variations in local feature preference across cortex can produce robust ensemble feature selectivity, when the information from many independent cortical regions is pooled in an optimized fashion. Here we investigated if ensemble patterns of fMRI activity in human visual areas contain robust information about motion direction. We performed conventional fMRI scans (3T scanner, voxel size 3x3x3mm) while subjects viewed random dots moving in 1 of 8 directions in each 16s stimulus trial. A linear decoder was trained to classify fMRI activity patterns induced by different motion directions by optimizing the voxel weights using linear support vector machines. Then, the decoder was evaluated with independe...