Accurate decomposition of standard MLP classification responses into symbolic rules

Abstract

In this work we determine hyper-plane equations from three MLP models. The first one is the standard MLP model, the second one is called OMLP (oblique MLP) and the last one is called IMLP (Interpretable MLP). From OMLP and IMLP, hyper-plane equations are determined easily, whereas for MLP we just give a sufficient condition for the detection of potential hyper-plane discriminators.

Our goal is to justify MLP classification responses in terms of symbolic rules. For this, we use a standard MLP network for classification and an IMLP network for justification of MLP responses. The system consists in the training of the IMLP network with MLP responses and in the extraction of symbolic rules from IMLP. The approach is sufficiently general to work even when input variables are continuous. Moreover, the justification provided by IMLP is accurate because if MLP and IMLP responses are contradictory with respect to a new unknown example, IMLP is retrained with the addition of the new example until the system becomes coherent. Finally, we show results given by a medical diagnosis application with continuous input variables.