Computing with Neural Networks: A Comparison of the Multi-Layer Perceptron and the Hierarchical Pattern Recognition Network for Classification Problems.

Abstract

Two artificial neural networks are compared in the realm of pattern recognition problems: the Multi-Layer Perception (MLP) using the Back Propagation algorithm for training and the Hierarchical Pattern Recognition Network (HPR). Both systems' architecture, node characteristics, and training algorithms are fully developed. Their decision boundary regions are demonstrated graphically on a series of 2 dimensional 2-class problems. The performance of both networks is tested on the Oxford Alvey Vowel Database and LANL's Stamped Metal Database. Additionally, LeCun's theory of Optimal Brain Damage is applied to the MLP in an attempt to reduce network size. The two case studies demonstrated that both networks perform equally well on real-world applications. The HPR network however, proved to be more advantageous in the realm of practical issues such as processing time, adaptability, memory requirements, and the ability to properly identify non-class data. LeCun's method of Optimal Brain Damage was effective in reducing the network size of the MLP and thus decreasing processing time as well as improving generalization performance.

Document Details

Document Type
Technical Report
Publication Date
Jan 01, 1990
Accession Number
ADA231584

Entities

People

  • Stacey W. Christian

Organizations

  • Air Force Institute of Technology

Tags

DTIC Thesaurus Topics

  • Algorithms
  • Boundaries
  • Brain Injuries
  • Case Studies
  • Classification
  • Databases
  • Neural Networks
  • Pattern Recognition
  • Perception
  • Recognition
  • Signal Processing
  • Training
  • Two Dimensional

Fields of Study

  • Computer science

Readers

  • Neural Network Machine Learning.
  • Systems Analysis and Design

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks