Analytical decision boundary feature extraction for neural networks with multiple hidden layers

Jinwook Go, Chulhee Lee

Research output: Contribution to journalArticle

Abstract

A feature extraction method based on decision boundaries has been proposed for neural networks. The method is based on the fact that normal vectors to the decision boundary provide the information necessary for discriminating between classes. However, it is observed that the previous implementation of numerical approximation of the gradient has resulted in some performance loss and a long processing time. In this paper, we propose a new method to calculate normal vectors analytically for neural networks with multiple hidden layers. Experiments showed noticeable improvements in performance and speed.

Original languageEnglish
Pages (from-to)579-587
Number of pages9
JournalLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume2756
Publication statusPublished - 2003 Dec 1

Fingerprint

Feature Extraction
Feature extraction
Normal vector
Neural Networks
Neural networks
Numerical Approximation
Processing
Gradient
Calculate
Necessary
Experiments
Experiment
Class

All Science Journal Classification (ASJC) codes

  • Computer Science(all)
  • Biochemistry, Genetics and Molecular Biology(all)
  • Theoretical Computer Science

Cite this

@article{e72ed044fd4d482cb996d2027cc661e7,
title = "Analytical decision boundary feature extraction for neural networks with multiple hidden layers",
abstract = "A feature extraction method based on decision boundaries has been proposed for neural networks. The method is based on the fact that normal vectors to the decision boundary provide the information necessary for discriminating between classes. However, it is observed that the previous implementation of numerical approximation of the gradient has resulted in some performance loss and a long processing time. In this paper, we propose a new method to calculate normal vectors analytically for neural networks with multiple hidden layers. Experiments showed noticeable improvements in performance and speed.",
author = "Jinwook Go and Chulhee Lee",
year = "2003",
month = "12",
day = "1",
language = "English",
volume = "2756",
pages = "579--587",
journal = "Lecture Notes in Computer Science",
issn = "0302-9743",
publisher = "Springer Verlag",

}

TY - JOUR

T1 - Analytical decision boundary feature extraction for neural networks with multiple hidden layers

AU - Go, Jinwook

AU - Lee, Chulhee

PY - 2003/12/1

Y1 - 2003/12/1

N2 - A feature extraction method based on decision boundaries has been proposed for neural networks. The method is based on the fact that normal vectors to the decision boundary provide the information necessary for discriminating between classes. However, it is observed that the previous implementation of numerical approximation of the gradient has resulted in some performance loss and a long processing time. In this paper, we propose a new method to calculate normal vectors analytically for neural networks with multiple hidden layers. Experiments showed noticeable improvements in performance and speed.

AB - A feature extraction method based on decision boundaries has been proposed for neural networks. The method is based on the fact that normal vectors to the decision boundary provide the information necessary for discriminating between classes. However, it is observed that the previous implementation of numerical approximation of the gradient has resulted in some performance loss and a long processing time. In this paper, we propose a new method to calculate normal vectors analytically for neural networks with multiple hidden layers. Experiments showed noticeable improvements in performance and speed.

UR - http://www.scopus.com/inward/record.url?scp=35248871580&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=35248871580&partnerID=8YFLogxK

M3 - Article

VL - 2756

SP - 579

EP - 587

JO - Lecture Notes in Computer Science

JF - Lecture Notes in Computer Science

SN - 0302-9743

ER -