Classification-based objective functions
Identifieur interne : 001176 ( Main/Exploration ); précédent : 001175; suivant : 001177Classification-based objective functions
Auteurs : Michael Rimer [États-Unis] ; Tony Martinez [États-Unis]Source :
- Machine learning [ 0885-6125 ] ; 2006.
Descripteurs français
- Pascal (Inist)
- Classification forme, Fonction objectif, Rétropropagation, Intelligence artificielle, Base donnée très grande, Reconnaissance caractère, Reconnaissance optique caractère, Base donnée, Saturation, Algorithme rétropropagation, Algorithme apprentissage, Réseau neuronal, Minimisation, Méthode heuristique, Optimisation.
- Wicri :
- topic : Intelligence artificielle, Base de données.
English descriptors
- KwdEn :
Abstract
Backpropagation, similar to most learning algorithms that can form complex decision surfaces, is prone to overfitting. This work presents classification-based objective functions, an approach to training artificial neural networks on classification problems. Classification-based learning attempts to guide the network directly to correct pattern classification rather than using common error minimization heuristics, such as sum-squared error (SSE) and cross-entropy (CE), that do not explicitly minimize classification error. CB1 iss presented here as a novel objective function for learning classification problems. It seeks to directly minimize classification error by backpropagating error only on misclassified patterns from culprit output nodes. CB 1 discourages weight saturation and overfitting and achieves higher accuracy on classification problems than optimizing SSE or CE. Experiments on a large OCR data set have shown CB 1 to significantly increase generalization accuracy over SSE or CE optimization, from 97.86% and 98.10%, respectively, to 99.11%. Comparable results are achieved over several data sets from the UC Irvine Machine Learning Database Repository, with an average increase in accuracy from 90.7% and 91.3% using optimized SSE and CE networks, respectively, to 92.1% for CB1. Analysis indicates that CB1 performs a fundamentally different search of the feature space than optimizing SSE or CE and produces significantly different solutions.
Affiliations:
Links toward previous steps (curation, corpus...)
- to stream PascalFrancis, to step Corpus: 000385
- to stream PascalFrancis, to step Curation: 000401
- to stream PascalFrancis, to step Checkpoint: 000346
- to stream Main, to step Merge: 001207
- to stream Main, to step Curation: 001176
Le document en format XML
<record><TEI><teiHeader><fileDesc><titleStmt><title xml:lang="en" level="a">Classification-based objective functions</title>
<author><name sortKey="Rimer, Michael" sort="Rimer, Michael" uniqKey="Rimer M" first="Michael" last="Rimer">Michael Rimer</name>
<affiliation wicri:level="2"><inist:fA14 i1="01"><s1>Computer Science Department, Brigham Young University</s1>
<s2>Provo, UT 84602</s2>
<s3>USA</s3>
<sZ>1 aut.</sZ>
<sZ>2 aut.</sZ>
</inist:fA14>
<country>États-Unis</country>
<placeName><region type="state">Utah</region>
</placeName>
</affiliation>
</author>
<author><name sortKey="Martinez, Tony" sort="Martinez, Tony" uniqKey="Martinez T" first="Tony" last="Martinez">Tony Martinez</name>
<affiliation wicri:level="2"><inist:fA14 i1="01"><s1>Computer Science Department, Brigham Young University</s1>
<s2>Provo, UT 84602</s2>
<s3>USA</s3>
<sZ>1 aut.</sZ>
<sZ>2 aut.</sZ>
</inist:fA14>
<country>États-Unis</country>
<placeName><region type="state">Utah</region>
</placeName>
</affiliation>
</author>
</titleStmt>
<publicationStmt><idno type="wicri:source">INIST</idno>
<idno type="inist">06-0297554</idno>
<date when="2006">2006</date>
<idno type="stanalyst">PASCAL 06-0297554 INIST</idno>
<idno type="RBID">Pascal:06-0297554</idno>
<idno type="wicri:Area/PascalFrancis/Corpus">000385</idno>
<idno type="wicri:Area/PascalFrancis/Curation">000401</idno>
<idno type="wicri:Area/PascalFrancis/Checkpoint">000346</idno>
<idno type="wicri:doubleKey">0885-6125:2006:Rimer M:classification:based:objective</idno>
<idno type="wicri:Area/Main/Merge">001207</idno>
<idno type="wicri:Area/Main/Curation">001176</idno>
<idno type="wicri:Area/Main/Exploration">001176</idno>
</publicationStmt>
<sourceDesc><biblStruct><analytic><title xml:lang="en" level="a">Classification-based objective functions</title>
<author><name sortKey="Rimer, Michael" sort="Rimer, Michael" uniqKey="Rimer M" first="Michael" last="Rimer">Michael Rimer</name>
<affiliation wicri:level="2"><inist:fA14 i1="01"><s1>Computer Science Department, Brigham Young University</s1>
<s2>Provo, UT 84602</s2>
<s3>USA</s3>
<sZ>1 aut.</sZ>
<sZ>2 aut.</sZ>
</inist:fA14>
<country>États-Unis</country>
<placeName><region type="state">Utah</region>
</placeName>
</affiliation>
</author>
<author><name sortKey="Martinez, Tony" sort="Martinez, Tony" uniqKey="Martinez T" first="Tony" last="Martinez">Tony Martinez</name>
<affiliation wicri:level="2"><inist:fA14 i1="01"><s1>Computer Science Department, Brigham Young University</s1>
<s2>Provo, UT 84602</s2>
<s3>USA</s3>
<sZ>1 aut.</sZ>
<sZ>2 aut.</sZ>
</inist:fA14>
<country>États-Unis</country>
<placeName><region type="state">Utah</region>
</placeName>
</affiliation>
</author>
</analytic>
<series><title level="j" type="main">Machine learning</title>
<title level="j" type="abbreviated">Mach. learn.</title>
<idno type="ISSN">0885-6125</idno>
<imprint><date when="2006">2006</date>
</imprint>
</series>
</biblStruct>
</sourceDesc>
<seriesStmt><title level="j" type="main">Machine learning</title>
<title level="j" type="abbreviated">Mach. learn.</title>
<idno type="ISSN">0885-6125</idno>
</seriesStmt>
</fileDesc>
<profileDesc><textClass><keywords scheme="KwdEn" xml:lang="en"><term>Artificial intelligence</term>
<term>Backpropagation</term>
<term>Backpropagation algorithm</term>
<term>Character recognition</term>
<term>Database</term>
<term>Heuristic method</term>
<term>Learning algorithm</term>
<term>Minimization</term>
<term>Neural network</term>
<term>Objective function</term>
<term>Optical character recognition</term>
<term>Optimization</term>
<term>Pattern classification</term>
<term>Saturation</term>
<term>Very large databases</term>
</keywords>
<keywords scheme="Pascal" xml:lang="fr"><term>Classification forme</term>
<term>Fonction objectif</term>
<term>Rétropropagation</term>
<term>Intelligence artificielle</term>
<term>Base donnée très grande</term>
<term>Reconnaissance caractère</term>
<term>Reconnaissance optique caractère</term>
<term>Base donnée</term>
<term>Saturation</term>
<term>Algorithme rétropropagation</term>
<term>Algorithme apprentissage</term>
<term>Réseau neuronal</term>
<term>Minimisation</term>
<term>Méthode heuristique</term>
<term>Optimisation</term>
</keywords>
<keywords scheme="Wicri" type="topic" xml:lang="fr"><term>Intelligence artificielle</term>
<term>Base de données</term>
</keywords>
</textClass>
</profileDesc>
</teiHeader>
<front><div type="abstract" xml:lang="en">Backpropagation, similar to most learning algorithms that can form complex decision surfaces, is prone to overfitting. This work presents classification-based objective functions, an approach to training artificial neural networks on classification problems. Classification-based learning attempts to guide the network directly to correct pattern classification rather than using common error minimization heuristics, such as sum-squared error (SSE) and cross-entropy (CE), that do not explicitly minimize classification error. CB1 iss presented here as a novel objective function for learning classification problems. It seeks to directly minimize classification error by backpropagating error only on misclassified patterns from culprit output nodes. CB 1 discourages weight saturation and overfitting and achieves higher accuracy on classification problems than optimizing SSE or CE. Experiments on a large OCR data set have shown CB 1 to significantly increase generalization accuracy over SSE or CE optimization, from 97.86% and 98.10%, respectively, to 99.11%. Comparable results are achieved over several data sets from the UC Irvine Machine Learning Database Repository, with an average increase in accuracy from 90.7% and 91.3% using optimized SSE and CE networks, respectively, to 92.1% for CB1. Analysis indicates that CB1 performs a fundamentally different search of the feature space than optimizing SSE or CE and produces significantly different solutions.</div>
</front>
</TEI>
<affiliations><list><country><li>États-Unis</li>
</country>
<region><li>Utah</li>
</region>
</list>
<tree><country name="États-Unis"><region name="Utah"><name sortKey="Rimer, Michael" sort="Rimer, Michael" uniqKey="Rimer M" first="Michael" last="Rimer">Michael Rimer</name>
</region>
<name sortKey="Martinez, Tony" sort="Martinez, Tony" uniqKey="Martinez T" first="Tony" last="Martinez">Tony Martinez</name>
</country>
</tree>
</affiliations>
</record>
Pour manipuler ce document sous Unix (Dilib)
EXPLOR_STEP=$WICRI_ROOT/Ticri/CIDE/explor/OcrV1/Data/Main/Exploration
HfdSelect -h $EXPLOR_STEP/biblio.hfd -nk 001176 | SxmlIndent | more
Ou
HfdSelect -h $EXPLOR_AREA/Data/Main/Exploration/biblio.hfd -nk 001176 | SxmlIndent | more
Pour mettre un lien sur cette page dans le réseau Wicri
{{Explor lien |wiki= Ticri/CIDE |area= OcrV1 |flux= Main |étape= Exploration |type= RBID |clé= Pascal:06-0297554 |texte= Classification-based objective functions }}
This area was generated with Dilib version V0.6.32. |