European Conference on Visual Processing
July - 2001 - Turkey
 
Oscillations and Rank-Order Coding
 
Simon Thorpe, Vanrullen Rufin, Arnaud Delorme & Gautrais Jacques
 

Centre de Recherche Cerveau & Cognition, Toulouse, France

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

 
    High-level object and scene categorisation can be achieved using only 10-20 ms of processing in areas such as V1, meaning that few neurons will have time to emit more than one spike. One coding scheme compatible with such severe temporal constraints is Rank Order Coding which uses the order in which neurons fire to code saliency - early firing implies high saliency. Although the scheme works well with flashed images, it is vulnerable to spurious stimulus-induced order effects caused, for example, by moving stimuli. However, by using internally generated sub-threshold oscillations, order-based processing can be decoupled from sensory input. By progressively depolarising a population of cells, those with the most effective inputs will tend to fire first, while less well-driven cells will fire later or not at all. Support for such a hypothesis comes from work by Fries et el (Nature Neuroscience, 2001) that demonstrated correlated latency shifting in populations of cortical neurons. Here, we use large-scale simulations with SpikeNet to demonstrate how such a strategy could be used to develop systems capable of fast and reliable object recognition in dynamically changing situations. Using oscillations also allows several waves of processing to be performed for the same input, thus increasing reliability.