GROUPE DE TRAVAIL : RESEAUX DE NEURONES

Centre Tolbiac, 90, rue de Tolbiac, Paris 13

A l'occasion du séjour à Paris 1 du professeur Manuel Grana (Université de San Sebastian) en tant que professeur invité, nous vous convions à venir nombreux assister aux 3 conférences qu'il donnera

1) le Vendredi 21 Mai, à 10h30,

Salle des thèses (C-22-04, ascenseurs rouges, 22ème étage)

Processing MRI images with the SOM

2) le Vendredi 4 Juin, à 10h30,

Salle des thèses (C-22-04, ascenseurs rouges, 22ème étage)

Processing Hyperspectral images with the SOM

3) le Vendredi 11 Juin, à 10h30,

Salle C-22-04 bis, ascenseurs rouges, 22ème étage

On the relationship between SOM and other Competitive Neural Networks

A bientôt,

Bien amicalement,

Marie Cottrell et Jean-Claude Fort

RESUMES

Processing MRI images with the SOM

Vector Quantization may be applied as a a kind of Bayesian filter forrestoration and unsupervised segmentation. As a filter it can produce smoothing for noise removal that preserves boundaries with great accuracy.

As a classification mechanism for unsupervised segmentation. We propose the use of the Kohonen's SOM for fast and robust VQ design. For the

determination of the appropriate number of codebooks we apply the so-called Occam filter approach. We will present results on unsupervised and supervised segmentation of Magnetic Resonance Images. For the latter, the Bayesian VQ performs as a preprocessing step.

Processing Hyperspectral images with the SOM

Hyperspectral images are becoming common in remote sensing imagery. We will introduce them and their applications to geology, agriculture, law enforcement, ecology. The processing of hyperspectral images involves in some cases dimensional reduction and classification, either in a supervised

way or in an unsupervised way. We will present some of the tools used for these task. We will introduce the advantages and problems of the SOM when applied to hyperspectral images. We propose a validation procedure using hypothetical supervised classifications, to avoid the problem of the lack of ground-truth samples for validation.

On the relationship between SOM and other Competitive Neural Networks

Competitive Neural Networks share a common structure of the adaptive rule, that essentially is a modification of the Simple Competitive Learning Rule.

We will review some of the published competitive neural network proposed in the literature trying to put them in the same computational framework. This framework is independent of the model embodied by the objective function minimized by the Competitive Neural Network. We consider that all the Competitive Neural Networks can be applied as continuation methods to search for the global optima of the Euclidean distortion in a robust way.

The Competitive Neural Networks could then be compared analytically on the basis of their properties as continuation methods.