Implementing the two-level threshold logic networks with near optimal number of hidden nodes

This paper proposes a method for constructing a threshold logic network with the optimal number of hidden nodes for classifying binary patterns into 2 classes. A subtraction after addition mechanism that decompose the given pattern class into optimal number of monotonic pattern subclasses while satisfying the necessary condition for linear separability is proposed. Also a sufficient condition for a monotonic pattern class to be linearly separable using a separating plane function is proposed. The proposed sufficient condition for linear separability can also be used to compute the connection weights and threshold values directly. The experimental result on a parity problem shows that our method can be effectively used to construct threshold logic networks with the optimal number of hidden nodes.