Realistic network models of distributed processing in the leech

The distributed nature of information processing in the b ra~n presents a great challenge to systems neuroscience. Whether one considers the processing of sensory information or the control of motor responses, the output of the nervous system is a function of large populations of simultaneously active neurons. In all but the simplest reflexes, neurons are arranged in bewildering networks of parallel and feedback patliways, making a purely intuitive understanding of the system diff~cult if not impossible. Neural models and computer simulations are approaches to this complexity that could help us achieve a deeper understanding of distributed processing. Useful predictions of a model must be directly related to the experiments that are performed. Since much of the current physiological effort in systems neuroscience is devoted to making singleunit recordings, models are needed that predict the responses of individual neurons. However, models sufficiently detailed to predict responses of individual neurons require a very large number of parameters, including the type and distribution of voltage-sensitive channels, the anatomical position of synapses, and the sign and strength of synaptic connections. Since the value of such parameters is generally not known, one approach is to study the effect on the model of systematically varying each parameter over its physiological range. In general, however, this is not an efficient strategy because there are often too many parameters and their physiological range is often quite large. There are now a variety of network optimization algorithms that adjust parameters such as connection strengths in artificial neural networks (Hinton 1989). The algorithms adjust each parameter in the network to reduce the overall error in the performance of the network. These algorithms differ according to the type of error information available and the way it is used to change the parameters. Optimization techniques are being applied to network models of known biological circuits where the number of parameters is too large to be studied individually. The implications of such models for our understanding of biological networks has been difficult to establish because of the many simplifying assumptions involved in the first generation of neural network models