Using a Proxy-Oriented Genetic Algorithm to Find a Millisecond-Scale Model of the Hippocampus

The Levy model is a neural network model of the CA3 region of the hippocampus. Previous work with the Levy model has shown success in modeling such hippocampally dependent tasks as trace conditioning, configural learning, spatial navigation, and sequence learning. Learning these tasks require network-scale behavior over simulated time-scales of minutes or longer. Most simulations of the model use simple McCulloch-Pitts neurons operating at time-scales of 15-30 ms.
Replacing the McCulloch-Pitts neurons with Izhikevich neurons allows the model to demonstrate biologically plausible neuron-scale behavior over time-scales of 1 ms and shorter. However, reproducing the network-scale behavior shown using the simpler McCulloch-Pitts neurons becomes more complicated due to the increased number of interacting parameters.
A genetic algorithm is used to explore these interacting parameters. Since the fitness function requires running a simulation of the CA3 region of the hippocampus, a proxy fitness function is used that simulates less than one second of time in the hippocampal model rather than a complete multi-minute simulation. The full fitness function only needs to be evaluated for parameter settings that pass a threshold value for the proxy function. Using a proxy-oriented genetic algorithm, settings were for the extended Levy model so that it can operate at millisecond time scales, demonstrate neuron-scale plausible behavior, while still demonstrating trace conditioning acquisition.