Feasibility of Identity Vectors for use as subject verification and cohort retrieval of electroencephalograms
Christian R. Ward, Iyad Obeid · 2016
The success of Identity Vectors in speech recognition as a tool for subject verification, language detection, word recognition, and accent/dialect classification suggests the technique is a robust method of unsupervised learning on high dimensional data, such as electroencephalograms. Tests run on the PhysioNet EEG Motor Movement/Imagery corpus concerning the matching of subject specific trials showed an average verification of 99% for the 109 subject-trial tests. Further tests on the ability to cluster repeated subject-trials produced at least one matching subject-trial for 60% of the subjects. The driving component of the Identity Vector process is the creation of Universal Background Models derived from single dimension Gaussian mixtures of user defined sizes operating on cepstrum feature coefficients. Taken as a whole, the results of this work indicate that Identity Vectors can be effective at distinguishing between subjects and show promise when asked to generate cohorts of related data.