Topology synthesis networks: self-organization of structure and weight adjustment as a learning paradigm

Abstract This work reviews the recent history of the design of neurocomputing algorithms and discusses the shortcomings that motivated the design of newer algorithms. In particular, it is proposed that recent results in a variety of scientific fields can be brought to bear to address a number of limitations in the current second generation of non-linear fixed structure networks. A new method called Self Organizing Neural Network (SONN) algorithm is reviewed, and its performance compared with the Backpropagation algorithm (Generalized Delta Rule). Previously presented results of time series prediction and signal separation are reviewed here. The SONN is an algorithm that constructs its own network topology during training, which is shown to be much smaller than the BP network, faster to train, and free from the trial-and-error nesign that characterizes BP.