Neural network analyses of stochastic information: application to neurobiological data.

Simultaneous recordings from over 50 neural cells were obtained from the dragonfly ganglia. To explore the biological information processing strategies reflected therein, data analysis methods were designed for use with artificial neural networks (ANN). Most methods are degraded by different cell spike trains that vary in mean firing frequencies by well over an order of magnitude. Based on underlying cell physiology, the occurrence of each spike is likely to be a stochastic function. To overcome such degradation problems in ANN use, a gaussian spike train representation was synthesized for each cell using raw data. This representation retained the exact spiking times and provided a biologically plausible probabilistic value for the time of occurrence for each spike. A 3-layer, feed-forward, ANN was trained on these data using a gradient descent learning algorithm. The task was to predict the neural activity at time (t + 1) given the neural activity at time (t). Following training, the network sum-squared prediction error was less than 0.01. Further, the temporal reproduction of the neural firing patterns was corroborated. The results indicated that the ANN could accurately reproduce the neural firing patterns in both the spatial and temporal domain using the stochastic spike train data. Encoding parameters for the spike trains using synthesized gaussian representations were optimized. The "lesion" studies were performed to determine the contribution of each cell to ANN predictions. The capability to "fine tune" both the information representation of spike trains and the ANN architecture provides significant advantages in the analysis of biological information processing by neural cells.(ABSTRACT TRUNCATED AT 250 WORDS)