Optimized Interconnections in Probabilistic Self-Organizing Learning

This paper focuses on self-organization of a multi-layered feed forward artificial neural network structure. Both the selection of interconnecti ons among neurons and their optimum weights are studied. In this learning structure, the neurons are sparsely connected and dynamically adjust their connectivity structure. Only the feed-forward propagation is used and each neuron dynamically adjusts its threshold based on the incoming data. By analogy to the signal weighting, this paper derived how to set the optimal interconnection weights for neuron's inputs. The binary input weight selection, suitable for hardware implementation, is discussed. Comparison between the binary and optimal weighting scheme is presented. Simulation examples for financial data analysis and power quality disturbance classification problems show the effectiveness of the proposed scheme.