UMR 5672

logo de l'ENS de Lyon
logo du CNRS
You are here: Home / Seminars / Machine Learning and Signal Processing / Feature-space selection in voxelwise encoding models with banded ridge regression

Feature-space selection in voxelwise encoding models with banded ridge regression

Tom Dupré la Tour (post-doctorant, Univ. Berkeley, CA, USA)
When Feb 03, 2021
from 06:00 to 07:00
Contact Name Pierre Borgnat
Attendees Tom Dupré la Tour
Add event to calendar vCal
iCal

Title: Feature-space selection in voxelwise encoding models with banded ridge regression.

Abstract: Ridge regression is used extensively in neuroimaging encoding models to model brain activity. However, ridge regression fails to account for the relative scales of different feature spaces that can be used jointly in a model. Ridge regression can thus be extended to banded ridge regression and optimize a different regularization hyperparameter per feature space. Here, we argue that banded ridge regression also performs a soft feature-space selection, effectively ignoring non-predictive and redundant feature spaces. We show that this feature-space selection leads to better predictive performances and to better interpretability. We then discuss the link between banded ridge regression and other sparsity-inducing linear models. We also address the computational challenge of fitting banded ridge regression models on large numbers of targets and feature spaces. Finally, we apply banded ridge regression to disentangle features extracted from pre-trained deep neural networks.

The talk will be online, on https://lpensl.my.webex.com/

Room : Webconf3