** Next:** Small Scale Variation
** Up:** Spatio-temporal Noise Modelling
** Previous:** Scale of Variation

In general, the large-scale variation represents deterministic
trends across one or more of the 4 dimensions of the observed data
space. Previous
work (16,43) has shown large scale temporal
noise processes. We do not incorporate these processes into the
model, but instead remove the worst of them by using high-pass
filtering as a preprocessing step (i.e. in
equation 2 we set ). In this work the high
pass filtering cut-off is chosen as the one which removes as much of the
temporal low frequencies as possible without degrading the signal.
There is increasing evidence that there is large scale spatial structure
in FMRI data possibly attributable to neuronal networks which are not
locked to an imposed task (34). When looking for task related activation
these large scale spatial networks are effectively ``noise/confounds''.
Ideally, a spatio-temporal model
would attempt to model these confounds out. However, in this work
we do not attempt to do so. This is an important area for future
work.