Skip to Main content Skip to Navigation

Traitement visuel rapide de scènes naturelles chez le singe, l'homme et la machine : une vision qui va de l'avant...

Abstract : At the boundary between neuroscience and artificial intelligence, computational neuroscience attempts to understand the tremendous computational capacities of the brain, and especially the high image processing efficacy of the visual system. My work is split into an experimental part and a modeling one. In the experimental part, I assessed the reasons for the speed and accuracy of visual processing. Natural photographs, that could include an animal or not, are flashed very briefly (20-30 ms) to subjects who must respond by releasing a button whenever the image contains an animal. Macaque monkeys perform this task with high accuracy, and although somewhat less accurate than humans, they are considerably faster. Then, I try to constrain the task in order to tackle the role of both the intrinsic properties of images - color, luminance, number of animals, visible parts of their body, species... - and their extrinsic properties - sequence effects, familiarity of images, instructions given to the subjects... Although certain conditions reduce the mean behavioral reaction times (RT), the fastest RTs, and EEG recordings that correspond to image processing, are left unaffected. Moreover, the fastest RTs do not seem to depend on specific images that would have been processed faster. Thus, this argues in favor of a rapid massively parallel visual processing, mainly automatic, where, because of the temporal constraints, each neuron hardly has time to fire more than once. Based on these constraints and those imposed by real neuron behavior, visual system structure and connectivity, I built a neural simulator - SpikeNET - to perform biologically plausible models that goes from basic visual processing (oriented bar selectivity) up to highly sophisticated visual processing (face recognition in natural images). From an image processing approach, the performance of these networks is impressive, especially because they rival with classical artificial intelligence techniques.
Document type :
Complete list of metadata
Contributor : Catherine Marlot <>
Submitted on : Thursday, June 8, 2006 - 10:55:40 AM
Last modification on : Friday, October 23, 2020 - 4:34:31 PM
Long-term archiving on: : Monday, April 5, 2010 - 10:29:46 PM


  • HAL Id : tel-00078924, version 1



Arnaud Delorme. Traitement visuel rapide de scènes naturelles chez le singe, l'homme et la machine : une vision qui va de l'avant.... Neurosciences [q-bio.NC]. Université Paul Sabatier - Toulouse III, 2000. Français. ⟨tel-00078924⟩



Record views


Files downloads