Distinguished Lecture Series: James Haxby (Princeton)
fMRl experiments produce large, numerically rich, but noisy data sets that pose a challenge for extracting the signal variance and establishing the correspondence between that signal and cognitive variables. Conventional analysis has reduced the dimensionality of fMRl data by searching for clusters of voxels that show similar responses to experimental manipulations and averaging the signal within those clusters. We have introduced a new approach to fMRl data analysis, “multi-voxel pattern analysis” of MVPA, that examines higher spatial frequency patterns of activity – the voxel-by-voxel variation of response within a region – and have shown that this method greatly increases the sensitivity of fMRI. These methods use pattern classifiers from machine learning such as linear discriminant analysis, neural networks, and support vector machines. In addition to increasing the sensitivity of fMRI analysis, MVPA can measure the multi-dimensional similarity structure of neural representations, which can be related to the multi-dimensional similarity structure of cognitive representations. The use of MVPA will be illustrated with studies of the representations of faces and objects in visual extrastriate cortex.