Knowledge extraction from SID epidemiological data using neural networks

Knowledge extraction is an important problem that has been little addressed by neural networks. In this work, we try to analyse the knowledge stored in a trained three-layer perceptron, using sudden infant death syndrome data. It is shown that when analysing the internal structure of the network, the classification solution realised by the network may be optimal in terms of classification results but not optimal in terms of knowledge representation. A simple method is proposed in order to reorganise and to extract knowledge stored in the synaptic weights