Parameter estimation using competitively inhibited neural networks
Michael D. Lemmon · 1990
Competitive interaction is a fundamental organizational principle of biological networks. This research investigates competition's role in network learning by focusing on a restricted class of laterally inhibited neural networks where the competition level has been parameterized. We call this network the competitively inhibited neural net (CINN). Analysis of the CINN reveals a simple methodology for predicting the network outputs and synaptic weights resulting from a constant input vector. This methodology is called the sliding threshold test. Its simplicity leads to an algorithmic characterization of the CINN which immediately suggests the use of fine-grained parallel computers for hosting these networks. The algorithm also suggests a conceptually simpler viewpoint of the network's dynamics which leads to a continuum model of CINN learning. This model is a nonlinear partial differential equation describing how the distribution of synaptic weights responds to a random sequence of constant inputs. Analysis of this model, using the method of characteristics, suggests that the CINN can be used to estimate the modes of unknown probability density functions and thereby identifies parameter (modal) estimation as an important application of the network. This research project shows how the CINN can be used as a parameter estimator in cases where the underlying observation operator may be nonlinear and the measurement noise has unknown a priori statistics. A demonstration of the CINN parameter estimator's performance is also presented.