Michel Salomon :"On the use of neural networks to solve problems"

Abstract :

Since the introduction of the perceptron at the end of the 1950s, neural networks have passed through several boom-and-bust episodes, becoming today the state-of-the-art approach for solving many classification and prediction problems. The backpropagation algorithm has thus given rise to the multilayer networks era. More recently, there has been a revival with the emergence of various deep neural networks also known as deep learning architectures. The research works described in this manuscript reflect on the one hand this network architecture evolution and on the other hand highlight the suitability of neural networks for successfully solving tasks in different areas. In a first part we show that a multilayer perceptron neural network can simulate lung motion being able to predict the trajectory of several feature points, that it can be a substitute for a computational fluid dynamics software for active airflow control, and finally allows to build also a true chaotic neural network. In the second part, convolutional neural networks (a main flavor of deep learning) are used to detect whether an image embeds a hidden message or not, the obtained steganalysis approach was at the time of its publication the state-of-the-art one. Then we propose a fully convolutional neural network for image denoising, this latter, which is an encoder-decoder, can deal with different kinds of noise. Finally, considering, the Reservoir Computing paradigm studied by collegues of our research institute, we study the application of Echo State Networks, a recurrent architecture, on a handwritten digits recognition problem (namely the well-known MNIST benchmark problem).

Composition of the jury :

- Christophe Cerisara, rapporteur, Chargé de Recherche CNRS (HDR) au Loria;

- Christophe Garcia, rapporteur, Professeur à l'INSA de Lyon;

- David Elizondo, rapporteur, Professeur à la De Montfort University, Leicester;

- Stéphane Chrétien, examinateur, Senior research scientist au National Physical Laboratory, London;

- Sylvain Contassot-Vivier, examinateur, Professeur à l'Université de Lorraine;

- Raphaël Couturier, examinateur, Professeur à l'Université de Bourgogne Franche-Comté.

Localisation : Amphi 3 du Bâtiment D - Génie Civil de l'IUT Belfort-Montbéliard, 19 avenue du Maréchal Juin à Belfort.