Dissertation/Thesis Abstract

Evolving spike neural network based spatio-temporal pattern classifiers with an application to identifying the alcoholic brain
by Roy, Arnab, Ph.D., State University of New York at Binghamton, 2014, 119; 3630956
Abstract (Summary)

We introduce a novel approach to evolving spike neural network (SNN) based Spatio-temporal (ST) pattern classifiers that can detect occurrences of hidden structures in a ST data. We test this learning paradigm to find characteristic electrical patterns in visually evoked response potentials (VERPs) generated by an alcoholic brain.

We cast the alcoholic classification task as a multiple feature selection (FS) problem. The FS problems are grouped under 2 classes: the spatial task and the temporal task. The objective of the spatial FS task is to choose a correct subset of electroencephalogram (EEG) leads (the spatial-features) along with the lead-weighs (numeric attributes) using which a composite signal can be created. The temporal FS task involves detecting temporal patterns that occur more frequently in the alcoholic composite signals than in the control signals. To facilitate the evolution of such a classifier, we introduce design rules for SNN based temporal pattern detectors (TPDs) and novel crossover operators for the simultaneous FS task.

The conventional techniques for characterizing the alcoholic VERPs use the information in the gamma-band (30 to 50 Hz) to develop a set of feature vectors and then train a classifier using these feature vectors. Using the SNN based evolutionary learning paradigm we were able to solve this problem in 1 step; the SNN performed both temporal feature extraction and classification. Unlike the conventional techniques we did not make any specific assumptions regarding the spectral characteristics of the data; we did not implement a gamma-band filter. Also, we located regions on the skull of an alcoholic subject that produced abnormal electrical activity compared to the controls. These regions are consistent with prior findings in the literature. The classification accuracy was measured as the area under the receiver operator characteristic curve (ROC). The area under the ROC curve for the training set varied from 90.32% to 98.83% and for the testing set it varied from 87.17% to 95.9%.

Indexing (document details)
Advisor: Schaffer, J. David, Laramee, Craig B.
Commitee: Lesperance, Leann, Sayama, Hiroki
School: State University of New York at Binghamton
Department: Bioengineering
School Location: United States -- New York
Source: DAI-B 75/11(E), Dissertation Abstracts International
Subjects: Biomedical engineering, Cognitive psychology, Computer science
Keywords: EEG, Evoked response potential, Evolutionary computation, Machine learning, Spike neural network, Subset selection
Publication Number: 3630956
ISBN: 9781321091748