The Interpretation Of Supervised Neural Networks

Abstract

Classij-ication of cancer and normal animal tissues is carried out on the basis of their 'H Nuclear Magnetic Resonance (NMR) spectra with neural networks trained by Back-Error Propagation (BEP), using two direrent costfunctions. A log-likelihood costfinction is shown to result in accurate out-of-sample generalisation with a smaller network than the usual Least Mean Squared (ZMS) error. ntejirst step in the interpretation of the operation of neural networks is to quantiJjr the relevance of the input parameters to the diagnosis of each tissue class. Two techniques for achieving this are investigated, namely the Jacobian method and a logarithmic sensitivity matrix. The latter is demonstrated to result in a clearer signature which is consistent across direrent network architectures and also broadly in agreement with conventional statistical correlations.