Supplementary Material Supplementary Methods and Results Multivariate Pattern Classification

A linear SVM is a discriminant function that attempts to fit a hyperplane that separates the examples from one experimental condition from another in a multivariate space with as many dimensions as voxels. When applied to fMRI data, a linear SVM fits a linear boundary that maximizes the distance between the most similar training examples from each experimental condition – these examples are referred to as the support vectors. The performance of the classifier is validated by evaluating the success of the classification boundary (defined using training) in predicting the experimental condition of previously unseen data examples. The hyperplane is determined to be orthogonal to the weight vector and is the direction in the data of maximum discrimination. The SVM weight vector is a linear combination, or weighted average, of the support vectors. That is, it is a linear combination of the volumes where responses to the two experimental conditions are most closely matched. Every voxel receives a weight, with larger weights indicating voxels of greater importance in defining the separating hyperplane. Given a positive and a negative label for each condition (+1 = Condition A;-1 = Condition B), a positive weight for a voxel means that there was a higher relative level of activity in that voxel for Condition A compared to Condition B in the support vectors, and a negative weight means activation was higher for Condition B (Mourao-Miranda et al. 2005). As the weights are defined by the support vectors, which in turn reflect the location of the separating hyperplane, both the sign (positive or negative) and size of the weights are defined multivariately. Furthermore, as SVM is a multivariate technique, the discrimination is based on the whole pattern. It is therefore not possible to make local inferences about single voxels-rather, weights of individual voxels should be interpreted within the context of their contribution to a wider discriminating pattern. Anatomical ROIs Figure S1 shows an example of the parcellation for the temporal ROIs in one participant. The average number of voxels across subjects in each temporal ROI was: left HG (77)