(+53) 7263-7100

The Parallel Searchlight MVPA Toolbox contains Matlab functions for classification and regression designed for parallelizing the computations of searchlight classifiers across the brain. This toolbox uses operations with sparse matrices with the purpose of reducing the computational time (CPU time). It avoids the sequential calculus of searchlight classifiers across voxels.

The maximum gain of computational speed is achieved for Gaussian Naive Bayes (GNB), since it is based on the assumption of independence between voxels. Nevertheless not all the algorithms improves the CPU time when compared with the sequential approach. This strategy produces a significant gain for those classifiers that requires iterative minimization of the cost function as: SVM and Logistic Regression (LR).
The implementations of SVM and Logistic Regression uses MEX compiled C code. The windows mex compiled functions are included in the toolbox. For other OS and the cpp files were also included.

This toolbox has been developed by Marlis Ontivero-Ortega, Agustin Lage-Castellanos, and Mitchell Valdés-Sosa from the Cuban Center for Neuroscience in collaboration with Giancarlo Valente and Rainer Goebel from the Cognitive Neuroscience Department at the Maastricht University.

The toolbox contains functions for the algorithms:

  • 1.    parallel_gnb.m and test_gnb.m: Parallel Gaussian Naive Bayes (GNB). Equivalent to linear discriminant analysis with the diagonal covariance assumption.
  • 2.    parallel_gd_svm.m and test_svm.m: Parallel Support Vector Machine (SVM) that minimize the primal objective function.
  • 3.    parallel_gd_lr.m and test_lr.m: Parallel Logistic Regression (LR). Equivalent to the ‘mnrfit’ Matlab function.
  • 4.    fMRI7T.mat:  Matlab Structure with the following fields:
  1.        data.X:                   [100x234067 double]                 features for MVPA analysis nsamples x nvoxels
  2.        data.SparseMat: [234067x234067 double]               sparse matrix that defines the voxels x searchlight relationship.
  3.        data.C:                 [100x1 double]                                   vector of labels ( 4 conditions)


A draft of the article with the mathematical formalism can be downloaded here. The matlab scripts for reproducing the figures presented in the article can also be downloaded.