Date of Award

5-31-1993

Document Type

Thesis

Degree Name

Master of Science in Computer Science - (M.S.)

Department

Computer and Information Science

First Advisor

Mark A. Gluck

Second Advisor

Catherine E. Myers

Third Advisor

Peter A. Ng

Abstract

Recent research exploring the use of neural networks for electro-encephalogram (EEG) pattern classification has found that a three-layer back-propagation network could be successfully trained to identify high voltage spike-and-wave spindle (HVS) patterns caused by epileptic seizures (Jando et. al., in press). However, there is no reason to predict that back-propagation is the best possible network architecture for EEG classification. A back-propagation neural network and a predictive autoencoder neural network were compared to determine which network was better at correct classifying both HVS and non-HVS patterns.

Both networks were able to classify 88%-89% of all patterns using a limited set of training data. The predictive autoencoder network trained with less epochs and appeared more resistant to overtraining. However, performance of the predictive autoencoder network may vary if it is stopped before it has trained for a sufficient number of epochs.

Share

COinS