search

wanderine / BROCCOLI

BROCCOLI: Software for Fast fMRI Analysis on Many-Core CPUs and GPUs
GNU General Public License v3.0
115 stars 38 forks source link

Elastic net regularization for GLM? #49

Open mmbannert opened 6 years ago

mmbannert commented 6 years ago

Dear Anders,

I was wondering if it would be feasible to implement elastic net regularization in GLM. It would be cool to estimate weights for very large design matrices with regularization at the speed of light :) Is there hope for this to come into existence?

Best, Michael

wanderine commented 6 years ago

Hi Michael,

to do a grid search for different L1 / L2 regularization settings is rather easy, but for first level fMRI there is temporal autocorrelation, which means that each voxel will have its own design matrix due to prewhitening. That part is therefore tricky to implement, and BROCCOLI currently does that partly on the CPU (4 iterations Cochrane-Orcutt).

2018-04-17 13:46 GMT+02:00 Michael Bannert notifications@github.com:

Dear Anders,

I was wondering if it would be feasible to implement elastic net regularization in GLM. It would be cool to estimate weights for very large design matrices with regularization at the speed of light :) Is there hope for this to come into existence?

Best, Michael

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/wanderine/BROCCOLI/issues/49, or mute the thread https://github.com/notifications/unsubscribe-auth/AEGryFZZqoDY-uB2mwZMFnkUPq5Z6gYlks5tpdYKgaJpZM4TYJNC .

-- Anders Eklund, PhD

mmbannert commented 6 years ago

Hey Anders, Right, I forgot about each voxel getting its own design matrix. Yes, it makes sense that this is not the highest priority for BROCCOLI. Thanks & best, Michael