[pymvpa] question about cross-subject analysis

[Rajeev Raizada]

> Yaroslav Halchenko wrote:
>
> I you only care about classification performances I bet you might soon
> hear from Raj...

Ah, well, I can't resist an invitation like that!  :-)

Hi John:
Definitely check out the hyper-alignment paper by Haxby et al.
which Yarik sent along earlier in the thread:
http://haxbylab.dartmouth.edu/publications/HGC+11.pdf
It's an important paper, and it shows a way of mapping
the voxel spaces of different subjects onto each other.

Andy Connolly and I recently published a paper
addressing a related but slightly different question,
namely how to map the neural-similarity space
of different subjects onto each other:
http://raizadalab.org/papers/RaizadaConnolly_JoCN_2012_uncorrected_proofs.pdf

Like you, we also used the Haxby 2001 data.
The fact that Jim and the PyMVPA developers
put that data online is a great service to the community.
Our Matlab and PyMVPA analysis code is here:
http://raizadalab.org/papers/RaizadaConnolly_SuppInfo_code.zip
and a brief description of how to run it is in this doc:
http://raizadalab.org/papers/RaizadaConnolly_JoCN_in_press_SuppInfo.pdf
The code has a lot of comments in it, so hopefully it's reasonably clear.

Our approach takes as its premise the idea
that the commonalities across subjects may emerge
from abstracting away from people's voxel-spaces,
as people's voxel-space "neural fingerprints" tend to be
somewhat subject-specific and idiosyncratic.
So, we abstract from voxel-space to similarity-space,
and then map people's similarity spaces onto each other
using a simple permutation-based approach.

The fact that the different subjects' VT-cortex masks
do not cover exactly the same voxels is thereby side-stepped,
as the analysis doesn't require the same voxels across subjects.
In fact, the neural similarity-spaces match up across subjs,
even when the sets of voxels that are used in each subj
are highly diverse (that's shown in Fig.3 of our paper).

I hope this helps.
I'd be very interested indeed to hear your thoughts.
Maybe try running our code on the Haxby data,
and see if it suggests any useful lines of attack.

Cheers,

Raj