Exploiting Don't-care Information in Neural Network Learning

In this paper, we present a novel neural network architecture called M-net, which exploits the don't-care information in training multilayer feedforward neural networks. Our method takes advantage of the user's prior knowledge as well as the neural network's ability to learn from examples. The user's prior knowledge is encoded in the form of don't-care inputs to reduce the number of training patterns required to represent a function. We derive the learning rule of M-net in the context of error backpropagation, and demonstrate its use on the priority decoding problem. Compared with conventional backpropagation networks , M-net drastically reduces the learning time while achieving superior quality.