Spatio-temporal associative memories-the role of inhibitory neurons in building synfire chains

Associative memories are of paramount importance for any system that must compute on real world data. Famous fixed point recurrent networks of the Hopfield type have shown their limits in both storage capacities and computational capabilities. Recent extensions to dynamic attractors seem to be able to increase both. A study of such a spatio-temporal memory model based on Abeles' synfire chains is presented. Experiments on compound words translation have revealed the promising computational properties of the model. It is shown here how, by controlling memory traces variability while learning, inhibitory neurons determine the number of representations that may be securely loaded in the system. This number is found to be proportional to the network size.