PREDICTION AND GENERALIZATION IN LOGICAL NEURAL NETS

A G-RAM is a logical neural network node with an in-built generalization capability. A method of estimating the generalization ability of G-RAM nets is proposed. An evaluation and comparison is made between G-RAMs and the conventional analogue feed-forward nets. The significant properties of the latter, such as the consistency condition and, more importantly, the relation between prediction and generalization are extended to G-RAMs. The method for estimating the generalization of the logical G-RAM net is verified with computer simulations.