Extension of Mutual Subspace Method for Low Dimensional Feature Projection

Face recognition algorithms based on mutual subspace methods (MSM) map segmented faces to single points on a feature manifold and then apply manifold learning techniques to classify the results. This paper proposes a generic extension to MSM for analysis of features in high-throughput recordings. We apply this method to analyze short duration overlapping waves in synthetic data and multielectrode brain recordings. We compare different feature space topologies and projection techniques, including MDS, ISOMAP and Laplacian eigenmaps. Overall we find that ISOMAP shows the least sensitivity to noise and provides the best association between distance in feature space and Euclidean distance in projection space. For non-noisy data, Laplacian eigenmaps show the least sensitivity to feature space topology.