archived @ http://www.sudoc.fr/074493981
La soutenance de ma thèse de doctorat de Sciences Cognitives a eu lieu le vendredi 7 février 2003 à 10h30, à l'ONERA, centre de Toulouse, 2 av. E. Belin, 31 Toulouse. The defence of my doctorate thesis in Cognitive Sciences took place February 7, 2003 at ONERA, Toulouse.
Le jury était composé de / The jury was composed of:
- Prof. Michel Imbert, Université Paul Sabatier,
- Prof. Jeanny Hérault, INPG, RAPPORTEUR
- Dr. Yves Burnod, Directeur de recherche INSERM, RAPPORTEUR
- Dr. Simon Thorpe, Directeur de recherche CNRS, CO-DIRECTEUR
- Prof. Manuel Samuelides, SUPAERO, DIRECTEUR
How to decipher vision's spiking code? Study of the parallel, asynchronous and sparse flow in ultra-rapid visual processing
We build and study dynamical models of visual coding as a parallel and asynchronous flow of information coded thanks to their exact succession in time. We will at first base the mechanisms of this code on the biological processes on the scale of the neurone and synapse. In particular, synaptic plasticity (Perrinet and Samuelides, 2000, Delorme and Al., 2001b; Perrinet et al. ., 2001) may induce the non-supervised extraction of coherent information in the flow of the neuronal impulses (Perrinet and Samuelides, 2002). Coding by the latency of the first spike can defines a code in the optic nerve based on multi-scale architecture. We extended these results by using an ecological approach allowing thanks to the statistics of natural images (Atick, 1992) the quantization of analog value by the spikes' rank. This visual code is based on a hierarchical feed-forward architecture (Van Rullen, 2001) which is distinguished, in addition to its simplicity, by its mathematical and computational performances. We will meet the needs for an effective model of Vision by defining a theory of over-complete event representation of the image (Perrinet and Samuelides, 2002). This formalization leads then to a strategy of a sparse spike code by defining lateral interactions (Perrinet and Samuelides, 2002c). Thanks to a reinforcement learning rule, this strategy can then be extended to a model of an adaptive cortical column which shows emergence of representation dictionaries, like those of Olshausen and Field (1998). Moreover, this paradigm adapts particularly to the construction of a saliency map. These techniques allow emergence of new tools for image processing and active vision which are particularly adapted to distributed computing architectures (Perrinet and al., 2003).
- Laurent Perrinet. Comment déchiffrer le code impulsionnel de la vision ? Étude du flux parallèle, asynchrone et épars dans le traitement visuel ultra-rapide, URL . 2003 abstract