Figure 2.

Overview of the training pipeline: the first session of the train set is vectorized and stacked into the matrix G_{1, train}. In each iteration, f features are selected from the vectorized train set using leverage score sampling. Then, the feature space of the vectorized validation set (a subset of the train set) matrices is restricted to the selected subset of features. The correlation between pairs of columns of the sub-sampled validation matrix is used to predict identity across sessions. In each iteration, the size of the feature set is incremented upto a maximum of 100. The optimal feature set is the one with maximum prediction accuracy. This feature set is then used in our experiments to predict identity of subjects in the test set.