Granular weights in a neural network
Scott Dick, Abraham Kandel · 2002
We investigate a mechanism for storing granular information in a neural network. The classic backpropagation network, which uses numeric connection weights, is known to be a universal approximator. However, training times for a backpropagation network can be very long, and the knowledge stored in these networks is exceptionally difficult to understand. Using granular weights can speed up network training and increase clarity, at the cost of some loss in accuracy. We describe our network architecture, called the granular neural network (GNN), which uses linguistic weights, as well as the rules of linguistic arithmetic, which were developed for this network. In an initial experiment, we found that the GNN completed training in an average of less than one tenth the number of epochs required by a backpropagation network on the Iris data set, when using a coarse granulation.