MODELING OF THE BONHOEFFER-EFFECT DURING LTP LEARNING

We investigate the effect of non-local synaptic enhancement within the framework of networks of formal neurons. Enhancement of nearby synaps es on a common axon without simultaneous post- and presynaptic stimula tion is modelled by a generalized Hebbian rule. The imprinted patterns of activity are no longer attractors of the network dynamics, but are processed instead. The network acts as a low band-pass filter for spa tially uncorrelated patterns and can discriminate spatially correlated patterns, according to their degree and range of correlation. While t he network is processing information, it still maintains some of the e rror-correcting properties of attractor neural networks.