Premium
Interactive neural‐network simulation: A textbook pattern‐recognition problem yields to art‐enhanced counterpropagation
Author(s) -
Korn Granino A.
Publication year - 1995
Publication title -
computer applications in engineering education
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.478
H-Index - 29
eISSN - 1099-0542
pISSN - 1061-3773
DOI - 10.1002/cae.6180030203
Subject(s) - computer science , artificial neural network , competitive learning , artificial intelligence , layer (electronics) , adaptive resonance theory , software , pattern recognition (psychology) , machine learning , chemistry , organic chemistry , programming language
We explain our motivation for truly interactive dynamic‐system simulation and apply DESIRE/NEUNET software to a classical textbook problem in neural‐network pattern recognition. Neural‐network designs which can distinguish two closely spaced spirals must discriminate precisely between complicated, nonconvex pattern clusters. This classical textbook problem is quite difficult. It was finally solved by an unconventional back‐propagation network using ad hoc layer skipping and, more accurately, by Boston University's ARTMAP algorithm. We discuss competitive learning and present a new accurate solution using a counterpropagation network whose novel competitive layer separates closely spaced patterns cleanly with the CLEARN algorithm, a very fast computer routine which emulates simplified adaptive‐resonance functionality.