Neural network optimization algorithms greatly enhance our ability to construct large-scale, dynamical models of highly interconnected networks. Until now, optimization has only been applied to networks of simplistic processing units, ignoring the integrative and temporal response properties of single neurons, thus limiting the predictive power of the models. The long-term goal of this project is to develop a hybrid modeling strategy in which optimization methods are applied to networks of realistic,multicompartmental model neurons. To accomplish this goal, we will construct a hybrid model of an actual distributed processing network composed of repeatably identifiable sensory, motor, and interneurons that computes a well-defined behavioral input-output function. Optimization will be used to predict the connectivity of as-yet-unidentified interneurons in the actual network and the predictions will be tested by identifying the interneurons by physiological and morphological means. Performance of the hybrid model will be assessed by comparing it to the performance of an a priori model in which all connection strengths are determined physiologically. The final model will be used to predict the loci of synaptic plasticity underlying nonassociative conditioning of the reflex by incorporating local learning rules and by optimization methods. The predictions will be tested by determining the actual plastic sites physiologically. This project will have the combined effect of enhancing the predictive power of optimized network models and illuminating the relation between computations at the single-neuron and network levels.
