> kentman234 Wrote:
> Although there are many things that must be further understood, in my opinion, the plan of the project should be more specified
> theoretically
This all depends on your research question. Are you interested in finding which brain regions can dissociate between different times of movies? In similarities of differences of time-courses in different regions? Look at individual differences, or commonalities? Do you want to go more data-driven or hypothesis driven?
> and also the practical aspects concerning AFNI i.e. the functions are to be used.
there's an AFNI matlab toolbox that supports I/O of AFNI briks. pyMVPA is recommended for machine learning of fMRI data.
-------------------------------------------------------
> nick Wrote:
> --------------------------------------------------
> -----
>
> > If you combine cross-validation with other
> > processing steps such as feature selection,
> please
> > consider that it's important to make sure that
> you
> > are not mixing up data used both for training
> and
> > testing. For example, you might bias your
> results
> > if you select informative features on the whole
> > dataset and then run cross-validation. The
> correct
> > approach is to run feature selection on the
> > training set in each cross-validation fold. See
> >
> [
miplab.unige.ch]
>
> > tti_brain_decoding_biases-ERRATA.pdf for
> details.
>
>
>
> I don't really understand why we shouldn't run the
> feature selection on the whole dataset. How could
> the results be biased?
Suppose you have a small set of brain volumes (samples) with only random noise of 50,000 voxels, and each volume has one of two classes (conditions) associated with it. Feature selection on the whole data-set may give you the 100 voxels that happen to be most informative for the condition of interest. Because these voxels are quite likely to be informative, at least to a certain extent, in most or all of your samples, that means that no matter how you do cross-validation with training and test sets, you are very likely to get classification performance above chance.
> Isn't it also possible to
> run the feature selection on the testing set?
You could, but that would also give you biased results (unless feature selection was done on the training set only - then it seems fine to use those selected features for the test set).
The issue above is known as "double dipping" and considered a Bad Thing.
To consider how Bad this Thing it is: respected researchers, much older than me, have had their apparently doubly dippy papers referenced to in red squares.
For more information, see:
Vul et al, "Puzzlingly high correlations in fMRI studies of emotion, personality, and social cognition"
Kriegeskorte et al, "Circular inference in neuroscience – the dangers of double dipping"
>
> >
> > Other than that I would suggest to start
> simple,
> > maybe with a region of interest that you expect
> to
> > be informative, or if you have no idea which
> area
> > is informative, consider using a searchlight
> > (information mapping).
> >
> > Hope this helps.
>
>
> You considered a simple way by starting with a ROI
> using a searchlight. How can I find this
> "information mapping" ?
google is your friend: search for "fmri information mapping" (without the quotes).
The first hit is Niko Kriegeskorte's paper in PNAS. On his website you'll find more resources, including a nice introduction on MVPA by Mur et al in SCAN.
Hope this helps,
Nick