Backpropagation for Energy-Efficient Neuromorphic Computing