Neural networks and the brain : associative learning and/or self-organisation ?

Experimental evidence suggests that modification of synaptic strength in the brain does not depend on co-activation of two connected neurons, as is assumed in most theoretical work since the proposals of Hebb (Hebb, 1949). Instead,through independent post-and presynaptic rules multiple modifications occur simultaneously at various sites in the nervous system. To account for this data, various researchers (Edelman, Fuster,...) propose an extension of the self-organising PDP approach to populational thinking. However, as in the PDP approach, the selection rules they propose only account for dynamical evolution of the system towards point attractors. The learning strategy of the networks is therefore still a purely bottom-up strategy. Experiments on visual perception seem to indicate that even low level visual processes can converge to more than one attractor (ambiguous figures, binocular rivalry), to limit cycles (oscillatory behaviour)or low-dimensional chaotic attractors. I argue to extend the neural network models of perceptual categorization to dynamical attractors and to include the multipticity of forms created by the autonomous, nonlinear brain dynamics as a complementary source of variation on which constraints of higher cognitive processes can act.