First Advisor

George G. Lendaris

Term of Graduation

Fall 1992

Date of Publication

10-14-1992

Document Type

Thesis

Degree Name

Master of Science (M.S.) in Electrical Engineering

Department

Electrical Engineering

Language

English

Subjects

Communication devices for people with disabilities, Gaze, Neural networks (Computer science), Visual evoked response

DOI

10.15760/etd.6142

Physical Description

1 online resource (3, ix, 107 pages)

Abstract

Severe motor disabilities can render a person almost completely incapable of communication. Nevertheless, in many cases, the sensory systems are intact and the eye movements are still under good control. In these cases, one can use a device such as the Brain Response Interface (BRI) to command a remote control (e.g. room temperature, bed position), a word-processor, a speech synthesizer, and so on.

The BRI is a gaze-addressing communication system that was developed by Dr. Erik E. Sutter at the Smith-Kettlewell Eye Research Institute, San-Francisco: a menu of communication objects (e.g. letter, word, command) is displayed on a TV screen; the disabled person selects an object by gazing at it; an electrode implanted in the visual cortex transmits the brain-wave response evoked by the visual stimulus associated with the object; a special device receives and analyzes the brain-wave signal to identify which object is being gazed at; and, finally, the action corresponding to the object is undertaken (e.g. change bed position). Currently, the brain-wave signal is analyzed by computing the correlation between the current brain-wave signal and previously recorded template signals. Although it was not part of the initial objective of this thesis, we found that normalizing the brain-wave signal before computing the correlation yields a significant improvement on the performance. This observation is of immediate utility since, to our understanding, the current implementation of the BRI does not normalize the brain-wave response.

This thesis explores using Artificial Neural Network (ANN) technology instead of the correlation technique to identify the brain-wave signals, with the important objective of shortening the response time. The method pursued was to effect correct classification on short portions of the brain-wave signals (short brain-wave signals lead to short response times).

In the present thesis, we conducted several experiments with two very different ANN architectures. In the course of the experiments, we repeatedly observed two phenomena that may have had an impact on the performance. We believe that these phenomena may happen in different problem contexts.

The first phenomenon, here called contrast-enhancement, refers to the fact that only the small deviations from the desired output values disappear during the training process whereas the larger ones do not. Associated with this phenomenon, we observed that the average output values of the ANN continue to get closer to the desired output values long after the (correct) identification decisions reached their maximum during the training process. The second phenomenon, here called nearby misclassification, refers to the fact that a misclassification is more likely to happen between two brain-wave signals evoked by "similar" visual stimuli. In this thesis, we describe a modification of the backpropagation paradigm aimed at improving the discrimination among arbitrarily defined classes of inputs (the error computed by the criterion function is made dependent on the class of the input).

Rights

In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

Comments

If you are the rightful copyright holder of this dissertation or thesis and wish to have it removed from the Open Access Collection, please submit a request to pdxscholar@pdx.edu and include clear identification of the work, preferably with URL.

Persistent Identifier

https://archives.pdx.edu/ds/psu/24665

Download

Share

COinS