Tight bounds on the size of neural networks for classification problems

Valeriu Beiu, Thierry De Pauw

Research output: Chapter in Book/Report/Conference proceedingConference contribution

6 Citations (Scopus)

Abstract

This paper relies on the entropy of a data-set (i.e., number-of-bits) to prove tight bounds on the size of neural networks solving a classification problem. First, based on a sequence of geometrical steps, we constructively compute an upper bound of O(mn) on the number-of-bits for a given dataset- here m is the number of examples and n is the number of dimensions (i.e., IR n). This result is used further in a nonconstructive way to bound the size of neural networks which correctly classify that data-set.

Original languageEnglish
Title of host publicationBiological and Artificial Computation
Subtitle of host publicationFrom Neuroscience to Technology - International Work-Conference on Artificial and Natural Neural Networks, IWANN 1997, Proceedings
PublisherSpringer Verlag
Pages743-752
Number of pages10
ISBN (Print)3540630473, 9783540630470
DOIs
Publication statusPublished - 1997
Externally publishedYes
Event4th International Work-Conference on Artificial and Natural Neural Networks, IWANN 1997 - Lanzarote, Canary Islands, Spain
Duration: Jun 4 1997Jun 6 1997

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume1240 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Other

Other4th International Work-Conference on Artificial and Natural Neural Networks, IWANN 1997
Country/TerritorySpain
CityLanzarote, Canary Islands
Period6/4/976/6/97

Keywords

  • Boolean functions
  • Classification problems
  • Entropy
  • Neural networks
  • Size complexity

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint

Dive into the research topics of 'Tight bounds on the size of neural networks for classification problems'. Together they form a unique fingerprint.

Cite this