Multiclass Sparse Bayesian Regression for fMRI-Based Prediction

Figure 6

Intersubject analysis. Maps of weights found by the different methods on the 2500 most relevant features by Anova. The map found by elastic net is difficult to interpret as the very few relevant features are scattered within the whole brain. SVR and VB-MCBR do not yield a sparse solution. Gibbs-MCBR, by performing an adaptive regularization, draws a compromise between the other approaches and yields a sparse solution, but also extracts small groups of relevant features.