... ICA¹

i.e. by feeding the results of the decomposition into the model-based higher-level analysis

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... subjects²

restricting ourselves to a simple single mean-group analysis for the moment.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... (PARAFAC³

Also known as 'canonical decomposition' (CANDECOMP [Carroll and Chang, 1970])

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... subjects⁴

Note that, similar to the two-dimensional case, we can freely pass scalar factors between estimates and also introduce permutations. Absolute amplitude in any of the factors is only meaningful when fixing all other factors to e.g. unit standard deviation or unit range.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... convergence⁵

e.g. when $\vert\vert\mbox{\protect\boldmath $A$}^{\mbox{\tiny new}} - \mbox{\protect\bold... ...w}} - \mbox{\protect\boldmath $C$}^{\mbox{\tiny old}}\vert\vert _F < \epsilon .$

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... model⁶

In its general form the multi-level GLM allows to combine contrasts of lower-level parameter estimates. Such designs, however, can be re-formulated as higher-level GLMs operating on simple parameter estimates; see [Beckmann et al., 2003a] for details.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... approximation)⁷

In the subject domain, the correlations with all other processes is not shown, as there are only 3 subjects in this study

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... map⁸

Identified as the map with highest spatial correlation with the GLM map out of 19 estimated sources; see figure 11(i). Also, the associated time course has highest correlation with the GLM regressor.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... normalised⁹

Estimated response size between sessions is normalised to unit standard deviation, thus showing relative response size only.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... map¹⁰

The model is ambiguous with respect to scalar factors and signs. The maps presented here have been scaled manually for comparison.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... solution¹¹

E.g. approximately 40min for the tensor-PICA estimation compared to ~8h for PARAFAC on the real FMRI data on a Compaq Alpha ES40 667MHz Server with Matlab 5.3 (excluding registration of individual session data to the template space) for the results presented in section 5.2.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.