To get greater efficiency, Google DeepMind’s researchers did what chip designers have long done, built a pipeline so that the learning rule for machine learning — backpropagation — is more efficient.