Hi Christian,
Thank you for the answer. I had read the article wrong, and thought that
tensor-ICA was applied to individual principle components, and not the
raw data. After reexamination of the article it is clear to me that the
dimension reduction is applied on all subjects enforcing identical time
loading vectors.
Thank you,
Gaël
On Sun, Jul 12, 2009 at 07:53:34AM +0100, Christian F. Beckmann wrote:
> Hi
> Tensor-ICA assumes consistent response in time by virtue of calculating
> the rank-1 approximation between the temporal response profiles of the
> different subjects/sessions. For resting FMRI data we recommend the
> concat-ICA approach (choose 'Multi-session temporal concatenation' in the
> GUI) The Tensor-ICA algorithm (incl the rank-1 approximation) is outlined
> in
> Beckmann and Smith. Tensorial extensions of independent component analysis
> for multisubject FMRI analysis. Neuroimage (2005) vol. 25 (1) pp. 294-311
> [1]http://dx.doi.org/10.1016/j.neuroimage.2004.10.043
> hth
> Christian
> On 12 Jul 2009, at 03:46, Gael Varoquaux wrote:
> Hi,
> I have been told that Tensor-ICA was not suited for paradigm-free group
> analysis, such as resting-state acquisitions. I was unable to understand
> the justification behind this. The claim was that for data acquired
> without paradigm on different subjects, the time series would be
> different, and thus the tensor model would not apply well. However, this
> puzzles me, as ICA is performed on principle components, and not raw
> time-series.
> Did I miss-understand the underlying reasons for avoiding Tensor-ICA on
> paradigm-less data? Is there a paper that would enlighten me?
> Cheers,
> Gaël
|