Sensory motor fmri

The data set is on

/NAS_II/Projects/MVPA_Language/motorsignal

How to run the data by yourself

Go to

/NAS_II/Projects/MVPA_Language/sparse_regularization/sm_sparse

Watch the tutorial video on how to run the elastic net environment on NAS_II (it's ready for you!!!):

Result files

The results are in

/NAS_II/Projects/MVPA_Language/motorsignal/sparse_LR/

There are 21 trials, and for each trial there are 21 observation.

There are 2 types of format

1) sweeping-across-parameter for each trial

trial${trial#}_bin_lr_all_21fold${fold_option}.mat

trial#: the number of trials, can be 1 to 21

fold_option: the assignment of train, validation, test. For example, 19:1:1 --> '1119'

The result is data structure containing lambda, alpha and its corresponding train, validation and test accuracy and the voxel consistency.

2) optimal parameter for each trial

trial${trial#}_${classif_reg_type}_cri0_bin_21fold${fold_option}.mat

classif_reg_type: the classifier and the regularization type, can be 'lr_none', 'lr_elnet', 'lr_lasso', 'lr_ridge'

The result report the optimal test accuracy by selecting the best validation accuracy from the result in 1).

Experiment#1: leave-one-obs-out ---- train:validation:test = 19:1:1

classifiers: LR+elastic net, lasso, ridge and no-regularization

The summary table is here

more details report:

1) Accuracy: accuracy of train set, validation set, test set

2) Consistency: the #overlapping voxels on different number of fold, i.e. there are V voxels that are selected in F folds.

The results suggest

1. elastic net regularization has the best accuracy over all trials
2. elastic net reg's consistency is better than lasso's
3. elastic net is more interpretable when compared with ridge or no-regularization. The size of selected features is not too big, consistent across folds and gives good accuracy