Rumelhart, Hinton, and Williams (1986) introduces the technique of the backpropagation of errors to train a Neural network.
This paper uses a feedforward network. However, it does not require that nodes are organized into strict layers; rather, it permits any directed acyclic topology.