Par Humberto Sossa, le mercrdedi 1er juin 2005 en salle C207 à 15 heures

Introduction

Humberto Sossa, hsossa@cic.ipn.mx, Centro de Investigación en Computación (CIC) del Instituto Politécnico Nacional (IPN), Mexico.

Abstract: An associative memory M is a kind of neural network with one layer. It can be seen as device by which one can associate patterns, input patterns xi with output patterns yi, with i=1,…,p. Each pair (xi,yi) is called in association. The whole set of associations is called fundamental set of patterns (FS). During trainig, with each association is build an associative matrix Mi. With the whole set of Mi’s is then build a final trainig memory M. During pattern recall a given key-pattern xj is presented to M and associated pattern yj is recalled. If the whole FS is recalled by M, we say that M provides perfect recall. For a noisy version xn of a given key pattern xk under some conditions M should also provide perfect recall of corresponding yk. In this talk we present a new model of associative memory by which we can recall patterns by means of key-patterns presented to a memory.

Applications go from 1) reconstruction of images contaminated, for example, with mixed noise. 2) object classification. 3) word recognition.