A model for temporal correlation of biological neuronal spike trains

Identification of temporal patterns of events is one of the objectives in real-time signal analysis. One of the techniques for identifying multiple temporal patterns in a noisy background environment, such as firing patterns of biological neurons, is the use of correlation analysis. A time-shifted hybrid model of a parallel-distributed processing network employing modified Hebbian rules and the backpropagation error-correction algorithm has been developed to compute the temporal correlation of firing patterns among many simultaneously recorded neuronal spike trains. These spike trains are analyzed by a multilayer network, with each hidden layer computing the correlation of time-shifted patterns of activity between the input and output. That is, once the network is trained to associate the input spike trains (time series) with the output spike train, the correlation among the input and output neurons at any lag time of k Delta t is represented by the connection weights at the kth hidden layer. These weights can be displayed graphically in a three-dimensional histogram form, providing an easily interpretable form of the input/output relationship. Simulation results show that the temporal correlation characteristics among different neurons are revealed by this technique.<>