Data representation for diagnostic neural networks

A paradigm for diagnostic neural network systems that emphasizes informative data representation and encoding and uses generic preprocessing techniques to extract knowledge from database records is discussed. The proposed diagnostic system differs from other approaches to automatic knowledge extraction in the following ways: by emphasizing the importance of intelligent encoding and preprocessing of raw data, rather than classifications; by demonstrating the importance of making a clear distinction between diagnostic and classification tasks; and by providing a generic, uniform representation for data records comprising interdependent, heterogeneous features. The correlation matrix memory (CMM), a linear system with a single-layer of input-output connections, that is used as the neural network system´s classifier is described. The limitations of the learning system are discussed.<>