Parallel distributed processing with multiple one-output back-propagation neural networks

A novel architecture of neural networks with distributed structures which is designed so that each class in the application has a one-output backpropagation subnetwork is presented. A novel architecture (one-net-one-class) can overcome the drawbacks of conventional backpropagation architectures which must be completely retrained whenever a class is added. This architecture features complete parallel distributed processing in that the network is comprised of subnetworks each of which is a single output two-layer backpropagation which can be trained and retrieved parallely and independently. The proposed architecture also enjoys rapid convergence in both the training phase and the retrieving phase.< >