Tree-based Neural Net (TBNN) for Learning "difficult" Concepts

Abstract: "This paper presents a new learning system and reports its first successful application to a medical domain. Polygraphic data were recorded during 8-hours sleep in 2 infants. A set of 15 parameters were estimated and submitted to an expert for classification. In this way, two data files were obtained and used as a test domain for the system. The tree-based neural net builds on the idea of generating a decision tree and translating it into a neural network architecture that is trainable by the backpropagation algorithm. The network has unambiguously defined topology and initial weights, and its subsequent training is very fast. The contribution of the system's individual aspects to the overall performance is studied. The achieved classification accuracy of cca 85% compares very favourably with other learning systems."