Discriminative Writer Adaptation

Abstract : We propose a general method for adapting a writerindependent classifier to an individual writer. We employ a mixture of experts formulation, where the classifiers are trained on weighted clusters of writers. The clusters are determined by which experts classify individual writing correctly. The method adapts by choosing the appropriate combination of classifiers for a new user. It applies to any probabilistic discriminative classifier, and adapts discriminatively without modeling the input feature distribution. We apply the method to online character recognition. Specifically, we use a mixture of neural networks as well as a mixture of logistic regressions. We train the mixture via conjugate gradient ascent or via the EM algorithm on 192,000 Latin characters of 98 classes and 216 writers, and show adaptation results for 21 writers.
Type de document :
Communication dans un congrès
Guy Lorette. Tenth International Workshop on Frontiers in Handwriting Recognition, Oct 2006, La Baule (France), Suvisoft, 2006
Liste complète des métadonnées

Littérature citée [12 références]  Voir  Masquer  Télécharger

https://hal.inria.fr/inria-00105168
Contributeur : Anne Jaigu <>
Soumis le : mardi 10 octobre 2006 - 14:54:16
Dernière modification le : mardi 10 octobre 2006 - 16:57:40
Document(s) archivé(s) le : mardi 6 avril 2010 - 19:11:52

Identifiants

  • HAL Id : inria-00105168, version 1

Collections

Citation

Martin Szummer, Christopher M. Bishop. Discriminative Writer Adaptation. Guy Lorette. Tenth International Workshop on Frontiers in Handwriting Recognition, Oct 2006, La Baule (France), Suvisoft, 2006. 〈inria-00105168〉

Partager

Métriques

Consultations de la notice

79

Téléchargements de fichiers

46