Synfire graphs: from spike patterns to automata of spiking neurons

Abstract

The concept of synfire chains has been proposed by Abeles as a reason-able biophysical model for cortical long-time correlations and replicating spike patterns in multi unit recordings. Some recent computational modelling approaches extend the model into a functional direction proposing that the synchronization of synfire chains may help to solve the binding problem of cortical information processing. In the present paper we investigate further computational aspects of synfire chains. First, we show how they can be used as spatio-temporal feature stores capable to learn, regenerate and recognize spatio-temporal signals. Thereby synfire chains introduce time into the static world of attractor neural networks as paradigms for cortical information processing. Then we extend the synfire chain model from linear autonomously evolving networks to graph-like structures with external input signals. Such synfire graphs can implement arbitrary deterministic and nondeterministic finite state automata. We prove formally that synfire graphs consisting of time-continuous spiking neurons can robustly process arbitrary long input words even if realistic postsynaptic potentials, bounded background noise and spike-timing jitter are taken into consideration. A single synfire node may consist of a single spiking neuron or a larger set of cells. In the latter case connections between two nodes can be diluted or have otherwise random synaptic efficacies. The extension of synfire chains to synfire graphs introduces operational (logical, procedural, cognitive) components into common modelling of Hebbian cell assemblies and brain functioning.