Proceedings of the Jacques Monod Conference
28 November-3 December - 1998 - Aussoua (France)
 
Learning and the Speed of Visual Processing
 
Arnaud Delorme, Michèle Fabre-Thorpe & Simon Thorpe
 

Centre de Recherche Cerveau & Cognition, Toulouse, France

arno@cerco.ups-tlse.fr, mft@cerco.ups-tlse.fr, thorpe@cerco.ups-tlse.fr

 
     Thorpe et al (1996) used event-related potentials to show that the human visual system can process complex natural images in roughly 150 ms, and recent behavioural work indicates that processing time may be even shorter in macaque monkeys (Fabre-Thorpe et al, 1998). Such data impose serious constraints on current models of visual processing, calling into doubt the classical view that information is transmitted using a firing rate code, since at each processing stage in the system, very few cells will be able to fire more than one spike before the next stage has to respond.
     We recently conducted a series of experiments based on the same visual categorisation paradigm in which subjects perform a go/no-go task, responding to a flashed photograph only when the picture contains an animal. However, whereas in the original experiments each image was presented only once (thus preventing the subjects from learning a particular processing strategy for each image), in these experiments subjects were trained for three weeks with a particular set of 200 images. At the end of the training period, the subjects were extensively tested with the same 200 familiar images mixed at random with 1200 totally novel images. A comparison of performance with the novel and familiar images revealed that although accuracy was somewhat higher for the familiar images, and the mean reaction time was somewhat shorter (424 ms for familiar stimuli, versus 444 ms for the novel ones), this improvement was essentially due to the elimination of long reaction time responses in the case of the familiar stimuli. We found no evidence that the fastest behavioural responses could be in any way speeded up by three weeks of training. This is clear from the fact that the 10th percentile point was at exactly the same latency, namely 360 ms, for both familiar and novel photographs. Furthermore, simultaneously recorded Event-Related-Potentials revealed that the strong differential response to targets and distractors which starts 150 ms after stimulus onset was no earlier with familiar than with novel stimuli.
     Such data suggest the existence of what we have termed "Ultra-Rapid Visual Categorisation", a mode of operation in which the visual system processes information so quickly that no further increases in processing speed can be obtained, even when extensive training is possible. We suggest that under such conditions, processing must be achieved using essentially feed-forward mechanisms and using a coding scheme that requires only one spike per neurone. Our recent work using SpikeNET, a program for simulating very large numbers of asynchronously firing integrate-and-fire neurones, indicates that such processing may indeed be possible, since we have successfully built a simple feed-forward architecture capable of accurately localising faces in natural images using only one spike per neurone (Van Rullen et al, 1998).
    
    Thorpe S., Fize D. & Marlot C. (1996). Speed of processing in the human visual system Nature, 381, 520-522.
    Fabre-Thorpe M., Richard G. & Thorpe S. J. (1998). Rapid categorization of natural images by rhesus monkeys
    NeuroReport, 9, 303-308.
    Van Rullen R., Gautrais J. & Thorpe S. J. (1998). Face detection using one spike per neurone Biosystems, (In
    press).