Average accuracy (across all 12 subject) as a function of the number of eigenvectors obtained with different dimensionality reduction techniques for the best classifier—k-NN (see also Table 3).

We included PCA, ISOMAP, LLE, Laplacian Eigenmaps, and t-SNE. The LDA dimensionality-reduction technique was not included in these curves because the maximal dimension for LDA is equal to the number of classes minus one.