Learning synergy in a multilevel neuronal architecture.

Abstract

An artificial worlds model of the brain has been developed that integrates memory, intraneuronal dynamics and multilevel evolutionary learning. The model includes two major subsystems. The first is a memory-manipulation scheme, called the reference neuron system, that serves to orchestrate a repertoire of neurons with different input-output capabilities. Signals impinging on these neurons are integrated by a cytoskeletal structure that is simulated as a cellular automation. The second subsystem is an evolutionary learning scheme, called the selection circuits system, that serves to train the neurons in the repertoire by varying the cytoskeletal proteins that control signal flow or readouts. The integrated system comprises two layers of cytoskeletally controlled neurons and two layers of reference neurons. Evolution can occur at the level of readout enzymes in neurons, at the level of proteins that control the flow of signals in the cytoskeleton and at the level of reference neurons that orchestrate the repertoire. The integrated system controls the motion of a modeled organism that is embedded in an artificial environment consisting of barriers, food and a target. The organism effectively learns to use patterns of barriers in its local environment to find the target, using food as a reward. Experiments with the model show that the integrated system enjoys significant computational synergies that make it more powerful than the component systems standing alone, that interactions between different levels at which variation can occur exert significant control over the tempo of evolution, that the synergies between different components and levels becomes more important as the environment becomes more complex and that mutation strategies that significantly slow down the rate of learning significantly decrease the degrading effects of environmental noise on performance.