Associative learning in evolved dynamical neural networks

At the behavioral level, learning is defined as a process that modifies the behavior of an organism in order to improve its performance. At the mechanistic level, learning may be implemented by a mechanism that modifies the strength of plastic synapses during learning. This implies that there is a sharp distinction between the dynamics of learning and the dynamics of the normal function of the network. Furthermore, the learning mechanism is assumed to have much slower dynamics than the dynamics of normal function. This dissertation explores the validity of these assumptions. At the behavioral level, we demonstrate that a real-valued genetic algorithm (GA) can successfully generate a neural network that can solve an associative learning task without synaptic plasticity. This result suggests that the sharp distinction between dynamics of learning and the dynamics of normal function is not necessary. Finite state automata (FSA) are used to analyze the dynamics of the non-plastic neural networks. The analysis shows that the phase portrait structure of the neural dynamics is important for the learning ability of the networks. The role of synaptic plasticity is also studied by using the GA to generate learning neural networks with synaptic plasticity. The study of these plastic networks suggests that the learning mechanism does not have to have slower dynamics than the dynamics of normal function. Rather, the plastic synapses act as extra neural states that rapidly change as part of the dynamics of the normal function of the network.