Elagage d'un perceptron multicouches : utilisation de l'analyse de la variance de la sensibilit\'e des param\`etres

TL;DR

This paper introduces a pruning algorithm for multilayer perceptrons that uses sensitivity analysis of parameter variance to optimize network structure, tested on simulations and compared with existing methods.

Contribution

It presents a novel pruning method based on variance sensitivity analysis, improving neural network structure determination.

Findings

Algorithm performs well on simulation examples

Compared favorably with three existing pruning algorithms

Demonstrates effectiveness of variance sensitivity analysis in network pruning

Abstract

The stucture determination of a neural network for the modelisation of a system remain the core of the problem. Within this framework, we propose a pruning algorithm of the network based on the use of the analysis of the sensitivity of the variance of all the parameters of the network. This algorithm will be tested on two examples of simulation and its performances will be compared with three other algorithms of pruning of the literature