5.3. FREM: fast ensembling of regularized models for robust decoding#
FREM uses an implicit spatial regularization through fast clustering and aggregates a high number of estimators trained on various splits of the training set, thus returning a very robust decoder at a lower computational cost than other spatially regularized methods. Its performance compared to usual classifiers was studied on several datasets in [Hoyos-Idrobo et al.].
5.3.1. FREM pipeline#
FREM pipeline averages the coefficients of many models, each trained on a different split of the training data. For each split:
aggregate similar voxels together to reduce the number of features (and the computational complexity of the decoding problem). ReNA algorithm is used at this step, usually to reduce by a 10 factor the number of voxels.
optional : apply feature selection, an univariate statistical test on clusters to keep only the ones most informative to predict variable of interest and further lower the problem complexity.
find the best hyper-parameter and memorize the coefficients of this model
Then this ensemble model is used for prediction, usually yielding better and more stable predictions than a unique model at no extra-cost. Also, the resulting coefficient maps obtained tend to be more structured.
There are two object to apply FREM in Nilearn:
nilearn.decoding.FREMClassifierto predict categories
nilearn.decoding.FREMRegressorto predict continuous values (age, gain / loss…)
They can use different type of models (l2-SVM, l1-SVM, Logistic, Ridge) through the parameter ‘estimator’.
5.3.2. Empirical comparisons#
126.96.36.199. Decoding performance increase on Haxby dataset#
In this example we showcase the use of FREM and the performance increase that it brings on this problem.
188.8.131.52. Spatial regularization of decoding maps on mixed gambles study#
The scikit-learn documentation has very detailed explanations on a large variety of estimators and machine learning techniques. To become better at decoding, you need to study it.
SpaceNet, a method promoting sparsity that can also give good brain decoding power and improved decoder maps when sparsity is important.