Download the non-parametric algorithm used to perform voxel decomposition.

The algorithm we used to infer components models each voxel as a weighted combination of canonical response profiles to the sound set. The algorithm searches for response profiles whose inferred weights across voxels are maximally non-Gaussian. Because summing independent variables causes them to become more Gaussian, non-Gaussianity can be used as a statistical signature to infer the original unmixed components. The algorithm is closely related to standard algorithms for independent component analysis (e.g. FastICA), but does not depend on any parametric assumptions about the type of non-Gaussian distribution that underlies the data. See the paper and the code for additional details.