Algorithm and implementation of a learning multiple-valued logic network

A learning technique and implementation for multiple-valued logic (MVL) networks are described. The learning problem is formulated as a minimization of an error function that represents a measure of distortion between actual and desired output. A gradient-based least-square-error minimization algorithm is used to minimize the error function, which in contrast to the backpropagation algorithm, does not involve a sigmoid function and requires only a simple sgn function in the learning rule. The algorithm trains the networks using examples and appears to be available in practice for most multiple-valued problems of interest. Circuit implementations of the learning MVL networks using CMOS current-mode circuits are described