Prevailing methods for analyzing functional MRI data are based on the General Linear Model (GLM). The popularity of the GLM is due to its linearity, which makes it theoretically and computationally tractable. In the framework of the GLM it is possible to model the BOLD response with a number of regressors in order to for example sensitize the detection of active brain voxels to an unknown delay in the BOLD response relative to stimulus onset. However, despite its name, the General Linear Model is not the most general linear technique; The GLM is a special case of a technique named Canonical Correlation Analysis (CCA). In the more general linear framework that CCA offers it is not only possible to account variations in BOLD response shape, but also variations in the shapes of active brain areas. This is important because optimal smoothing of fMRI images should be matched to the active areas. In the GLM, smoothing is carried out with a fixed Gaussian filter, but with CCA an adaptive smoothing is effectively implemented. In the first image below, an example of a spatial basis function set is shown. These basis functions model the spatial extent of an active brain area and they are used very much in the same way as the temporal basis functions used for modeling the temporal evolution of the BOLD response. In the last figure the result using fixed smoothing (i.e. GLM) and steerable smoothing (i.e. CCA) applied to a synthetic data set is shown. The superior detection capabilities of the CCA approach is quite evident.