Variable Time Delays and Representational Capacity in Sparsely Connected Populations of Spiking Neurons

Successive generations of artificial neural networks have leveraged their multiplicity of connections and weights for significant improvements in information processing capability and memory capacity. The most recent generation of artificial neural networks, third generation networks, consist of spiking neuron models that attempt to mimic the complex dynamic features exhibited by real biological neurons in the hopes of improvements in computational and representational capacities. While the theoretical capabilities of these networks are impressive, understanding the nature and extent of their computational advantages, and the appropriate network architectures and algorithms necessary for their successful exploitation, have lagged far behind the theory. With this in mind, we herein explore the representational capacity of two related forms of neural networks: synfire chains, and polychronic networks. We find that the computational capacity of such cellular assembly based networks increases with the size of between-neural-pool time delays and that for relatively small changes in time delay, linear increases in network representational capacities are obtained.