Thesis Analysis of Lvq in the Context of Spontaneous Eeg Signal Classification

OF THESIS ANALYSIS OF LVQ IN THE CONTEXT OF SPONTANEOUS EEG SIGNAL CLASSIFICATION Learning Vector Quantization (LVQ) has proven to be an e ective classi cation procedure. Since its introduction by Kohonen in 1990 several extensions to the basic algorithm have been proposed. This paper investigates what and how LVQ learns in the context of EEG signal classi cation. LVQ is shown to be comparable with other Neural Network algorithms for the task of classifying electroencephalograph (EEG) signals, yielding approximately 80% classi cation accuracy for three out of the four subjects tested when di erentiating between two di erent mental tasks. The best classi cation accuracy was obtained with unnormalized, sixth-order autoregressive, AR(6), coe cients derived from raw, un ltered EEG signals. The LVQ2.1 algorithm outperformed any of the other traditional LVQ algorithms tested, yielding a slightly higher classi cation accuracy than the LVQ3 algorithm. The highest classi cation accuracy for di erentiating between two tasks was obtained using 16 codebook or reference vectors per task and a learning rate, , of 0.1. The value of the window width parameter had no e ect on the classi cation accuracy of LVQ2.1 The window width parameter speci es the width of a window centered around the hyperplane separating the two reference vectors to be updated. Reference vector updates only take place if the data vector currently being considered lies in the area de ned by the window. Initializing reference vectors at random data points resulted in an insigni cantly higher classi cation accuracy than using K-means to initialize the reference vectors. Using the OLVQ algorithm as part of the initialization procedure did not a ect the overall