A Model for Programmability and Virtuality in Dynamical Neural Networks

In this dissertation a fixed-weight architecture for Continuous Time Recurrent Neural Networks (CTRNNs) is proposed in order to give an account for biological phenomena, controlled by neuronal activity, in which changes of behavior occur so fast that presumably no changes in the involved neuronal connectivity are possible. The proposed model possesses the following properties: a. the neural network variables have a direct biological interpretation; b. the change of behavior is controllable by auxiliary (programming) inputs; c. a single fixed-weight neural network has the capability to exhibit a wide repertoire of different behaviors given the appropriate auxiliary inputs. Such properties allow the model to be biologically plausible on the neural level and, at the same time, should sustain a programmability / virtuality capability usually associated only with the algorithmic, symbolic systems used in the high level functional modeling of biological systems. A number of experiments are performed which corroborate: 1) the capability of the proposed architecture to be programmed with auxiliary inputs in order to reproduce the dynamical behaviors of networks with weight values coded by the auxiliary input; 2) the robustness of the proposed architecture w.r.t. variations of the I/O time scales.