Backpropagation for Neural Dnf{ and Cnf{networks

The architecture of a neural network with its links and weights can be viewed as a knowledge representation. To overcome the black{box problem of not knowing what knowledge is hidden in a neural architecture, we present a strictly logically operating network. Each neuron represents either a disjunc-tion or conjunction of its inputs and the net thus performs the function of a logic formula. This formula can be extracted from the net after completed training. The well{known backpropagation algorithm is adapted to train such logical neural nets. An on{line pruning algorithm is also implemented to eliminate redundant weights.