Ensembling brain regions for brain decoding.

In this study, we propose a new method which ensembles the brain regions for brain decoding. The ensemble is generated by clustering the fMRI images recorded during an experimental set-up which measures the cognitive states associated to semantic categories. Initially, voxel clusters are formed by using hierarchical agglomerative clustering with correlation as the similarity metric. Then, for each voxel cluster, a support vector machine (SVM) classifier is trained to estimate the class-posteriori probabilities. Lastly, the class-posteriori probabilities are ensembled by concatenating them under the same feature space, which are then used to train a meta-layer SVM for the final classification of the cognitive states. By using the voxel clusters, we aim to utilize the distributed, but complementing nature of the semantic representations in the brain and improve the classification accuracy. Thus, we make an existential claim that the brain regions provide a natural basis for ensemble learning which should be superior to the random clusters formed over a selected set of voxels. Our approach yields to better classification accuracies in Mitchell dataset on most of the subjects, when compared to state-of-the-art which emphasizes voxel selection and ensemble learning with random subspaces.