[pymvpa] Searchlight sl_map = accuracy?

[Nynke van der Laan]

Thanks for your quick reply again! :)

I am still a little confused. In several of the other examples, the
default values for errorfx are used. For example in '9.3.5 The minimal
searchlight example', the following code is used to prepare the
analysis:
cv = CrossvalidatedTransferError(
     TransferError(LinearCSVMC())
     OddEvenSplitter())
In the example 9.3.6 Searchlight on fMRI data this is:
clf = LinearNuSVMC()
cv = CrossValidatedTransferError(
     TransferError(clf),
     NFoldSplitter())
If I see this correct, the only difference between the examples is the
classifier used. The results would be reported as the same measure
(i.e., the default errorfx).
However, in the first example, the best performing sphere is the
sphere with the highest error (the best performing sphere is retrieved
by max(sl_map). I would expect that the best performing sphere is the
one with the lowest error.... (i.e., the best sphere would be the one
with the highest accuracy and thus the lowest error)

I was wondering if there is a way to get a sl_map with the accuracies
instead of the errors? I couldn't find a proper errorfx argument for
that. Or should I than just do that myself by for each datapoint doing
1 minus the value I get now?

Thanks in advance!

Best,
Nynke


On Tue, 22 Feb 2011, Nynke van der Laan wrote:
> calculation). So the CrossvalidatedTransferError is the mean of the
> transfererrors (i.e., accuracies) over the validations. From your

yes

> reply I understand that the outcomes of the analysis in the example
> (the Searchlight on fMRI data) are not the accuracies but the inverse
> (i.e., 1 - accuracy). Would you please confirm if this is right?

I do confirm that with the default argument of errorfx for
TransferError, you would obtain errors, not accuracies, in the
corresponding Searchlight.

Are you getting strange results with low errors in the regions where
there could be no signal (e.g. outside of the brain)? ;-)