Self-organizing neural network implemented in analog circuitry requires supervisional, scaling and parametric optimization attributable to 1/f noise

Abstract

This thesis describes the utilization of an analog computer for the purely analog electronic implementation of an unsupervised neural network. Results indicate that the algorithmic principles embodied 1n the nonlinear ordinary differential equations of Grossberg's Outstar learning model are destabilized by the inescapable multiplicative biases of the analog circuit attributable to 1/f noise drift. A multiplicative offset perturbation model was developed, and demonstrated numerically to simulate the types of instabilities discovered in the analog implementation. Supervisional, scaling, and parametric design requirements enabling stable analog implementation were developed. The optimized circuit was demonstrated to perform training and reca 11 of spatial patterns with enormous speed, robustness, and gracefully degrading performance, even when the optimized requirements were modestly changed. We formally present a hypothesis called 11 0utstar Neural Networks in Natural Noise are Non-implementable" (or Outstar N5) as a generalization of these results. Several mathematical techniques are suggested to facilitate weaker convergence criteria derivation, in hopes of extending the strictly real-time self-organizing Outstar theory to contend with the destabilizing 1/f phenomena. These techniques include the fractional-order calculus, distribution theory, and Ito-Stratonovich stochastic calculus.

Notes

This item is only available in print in the UCF Libraries. If this is your thesis or dissertation, you can help us make it available online for use by researchers around the world by STARS for more information.

Graduation Date

1990

Semester

Fall

Advisor

Georgiopoulos, Michael

Degree

Master of Science (M.S.)

College

College of Engineering

Department

Electrical Engineering

Format

PDF

Pages

285 p.

Language

English

Length of Campus-only Access

None

Access Status

Masters Thesis (Open Access)

Identifier

DP0027719

Subjects

Dissertations, Academic -- Engineering; Engineering -- Dissertations, Academic

This document is currently not available here.

Share

COinS