21st European Conference on Visual Perception
August 24-28 - 1998 - Oxford (England)
 
Ultra-Rapid Visual Categorisation: Large-scale simulations using SpikeNET
 
Simon Thorpe, Rufin van Rullen, Arnaud Delorme & Jacques Gautrais.
 

Centre de Recherche Cerveau et Cognition (UMR 5549) Faculté de Médecine de Rangueil, 133 route de Narbonne, 31062 Toulouse, France.

thorpe@cerco.ups-tlse.fr , rufin@cerco.ups-tlse.fr, arno@cerco.ups-tlse.fr, gautrais@cerco.ups-tlse.fr

 
     The ability of both monkeys and humans to perform ultra-rapid visual categorisation of previously unseen natural scenes poses a number of problems for conventional views of visual processing. For example, we have argued that firing rate codes may not be fast enough to be compatible with the short latency of category specific evoked potentials. An alternative scheme uses the analog-to-delay transformation characteristic of retinal ganglion cells to produce an asynchronous wave of spikes in which information can be encoded by the order in which cells fire.
     To test the plausibility of this hypothesis we have performed large scale simulations of the propagation of spikes through the visual system using SpikeNET, a neural network simulator. SpikeNET is suitable for simulating networks involving millions of neurones and hundreds of millions of connections and can be used to study a wide range of architectures. We have also been working on hardware implementations of SpikeNET capable of supporting near real-time performance.
     We have found that even relatively simple feed-forward architectures can perform complex visual processing tasks that include the localisation of faces in natural scenes. One of the most remarkable features of these simulations is that they can be made to work even under conditions where each neurone in the network is only allowed to emit a maximum of one action potential, thus effectively excluding any possibility of using a conventional firing rate code.