HONEST: a new high order feedforward neural network

A frequently voiced complaint regarding neural networks is that it is difficult to interpret the results of training in a meaningful way. The HONEST network is a new feedforward high order neural network (HONN) which not only allows a fuller degree of adaptability in the form of the nonlinear mapping than the sigma-pi model, but has a structure that can make it easier to understand how the network inputs come to be mapped into the network outputs. This structure also makes it easier to use external expert knowledge of the domain to examine the validity of the HONEST network solution and possible to reject some solutions. This is particularly important for embedded, failure-critical systems such as life-support systems. We have applied the HONEST network to the problem of forecasting the onset of diabetes using eight physiological measurements and genetic factors. We obtained a successful classification rate of 83% compared to a 76% rate that had been obtained by previous researchers.