This work presents a systematic approach to parallel implementation of feedforward neural networks on an array of transputers. The emphasis is on back propagation learning and training-set parallelism.