IMS algorithm for learning representations in Boolean neural networks
Nupur Biswas, T.V.M.K. Murthy, Meera Chandrasekhar · 1991
A novel algorithm for learning representations in Boolean neural networks, where the inputs and outputs are binary bits, is presented. The algorithm has become feasible because of a newly discovered theorem which states that any nonlinearity separable Boolean function can be expressed as a convergent series of linearly separable functions connected by the logical OR (+) and the logical INHIBIT (-) operators. The formation of the series is carried out by many important properties exhibited by the implied minterm structure (IMS) of a linearly separable function. The learning algorithm produces the representation much faster than backpropagation and, unlike the latter, does not encounter the problem of local minima. It also successfully separates a linearly separable function and obtains the perceptron solution in the presence of a spoiler vector, a situation where backpropagation is guaranteed to fail.>