Partenaires

CNRS
Logo tutelle
Logo tutelle
Logo tutelle


Rechercher

Sur ce site

Sur le Web du CNRS


Accueil du site > Séminaires > Probabilités Statistiques et réseaux de neurones > Rule Extraction from Prototype-based Classifiers

Vendredi 25 mars 2005 à 10 heures

Rule Extraction from Prototype-based Classifiers

Barbara Hammer (Technical University of Clausthal, Allemagne)

Résumé : Neural methods constitute quite efficient classifiers, but they have the lack that their decisions can usually not be understood by humans since distributed, subsymbolic modelling is used instead of symbolic rules. This is a drawback if machine learning is used to acquire explicit knowledge to guide humans, if an explicit identification of the relevant parts for a decision is required, or safety-critical problems are tackled. Nevertheless, subsymbolic methods have the benefit to provide smooth, fault-tolerant, and robust solutions for real-life problems. This fact makes the development of intuitive hybrid methods which combine neural classifiers with explicit symbolic rules attractive. Within the talk, prototoype-based classifiers will be considered and two different ways to efficiently extract rules from the neural classifiers will be presented : the first method directly extracts decision trees based on trained network and, afterwards, optimizes the result with respect to the decision borders automatically.

The second method iteratively extracts a decision border and trains a neural classifier on the resulting new subproblems until a desired degree of accuracy is achieved. Both methods are thereby intuitive and require little extra cost with respect to the neural classifier. We demonstrate the good performance of the methods on several classical rule-extraction benchmark problems.

Voir en ligne : le site de Barbara Hammer

Dans la même rubrique :