Conference of neural Binding in space and time
March (15-18) - 2000 - Leibzig (Germany).
 
Object recognition using spiking neurons II: Spatial attention explained by temporal precedence of information.
 
Rufin van Rullen, Arnaud Delorme & Simon Thorpe
 

Centre de Recherche Cerveau et Cognition, CNRS, Université Paul Sabatier, 133 route de Narbonne, F-31062 Toulouse, France

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

 
    Recent models of object recognition based on the temporal properties of spiking neurons have proved highly efficient, even when processing complex natural images (see Delorme, Van Rullen & Thorpe, this conference). These models overcome the problem of binding location with identity by having one neuron selective for an object at every position in the visual field. However, it is well known that, due to the resource limitations of biological visual systems, real neurons have increasing receptive field sizes, covering most of the visual field in higher visual areas such as IT. This loss of spatial resolution makes such networks vulnerable to the binding problem, because features belonging to different objects can become mixed up when more than one object is present.
    Under such conditions the binding problem can still be avoided by using directed spatial attention. We propose a simple mechanism for spatial visual attention that involves selectively lowering the thresholds of neurons with receptive fields in the attended region. Information about this region of interest propagates faster, and reaches the neurons in higher visual areas earlier. Therefore the early part of these neurons' responses will only be affected by the attended stimulus. As a result, unattended information can be filtered out by desensitization or inhibition mechanisms. Our hypothesis only requires that the dynamical asynchronous properties of real neurons be taken into account. We implement this mechanism in a simple model of the visual cortex, and show that it accounts very well for both the need of computational efficiency and the limited resources of the system. A remarkable feature of the proposed model is its purely feed-forward organization. No loops, feedback or lateral inhibition are required to enhance information at the attended location, or to filter out unattended information. A variety of experimental results can be interpreted in the light of this hypothesis. Finally, the binding problem is overcome because the use of attention means that only the features coding for the attended stimulus will be activated.