Méziane Yacoub

Discriminative feature extraction and selection applied to face recognition

We propose an integrated approach to feature and architecture optimization for convolutional connectionist models. The goal is to select single features which are likely to have good discriminatory power and extract nonlinear combinations of features with the same aim. In particular, the focus is on the interaction of the feature extraction and selection modules with the recognizer design. We propose a pruning-based method called /spl epsi/HVS (extended HVS), where the use of a priori knowledge is adaptively optimized during a discrimination training criterion aiming at minimum classification error. Results demonstrate the selection approachs effectiveness in identifying reduced architectures with the same recognition accuracy.

Architecture optimization in feedforward connectionist models

Given a set of training examples, determining the number of free parameters is a fundamental problem in neural network modeling. The number of such parameters influence the quality of the solution obtained. This paper deals with the problem of adapting the effective network complexity to the information contained in the training data set, and the task’s difficulty. The method we propose consists of choosing an oversized network architecture, training it until it is assumed to be close to a training error minimum then selecting the most important input variables and pruning irrelevant hidden neurones. This method is an extension of our previous one used for input variables selection, it is simple, cheap and effective. We show its effect experimentally through one classification and one regression problem.

Une mesure de pertinence pour la sélection de variables dans les perceptrons multicouches

Ce papier est consacre essentiellement a notre mesure heuristique, nommee HVS (Heuristique for Variable Selection)[YAC 97], que nous utiliserons pour la selection de variables. HVS ne demande que peu de calculs simples, faciles a implementer. Nous testerons son efficacite sur un probleme de discrimination et un probleme de regression, apres avoir montre sa capacite de detection et de quantification de pertinence.

A neural network methodology for machines' class identification

Before learning a given machine coded by a set of input-output pair sequences, we are interested in identifying whether this machine is a deterministic finite state machine, and if so whether it is a definite memory machine, a finite memory machine, or has an infinite order. If the result is that it has a finite memory order, we attempt to approximate its input and output memory order. A methodology is proposed, and experiments on different machines are presented.