Since the dawn of the modern digital era, scientists and engineers have taken inspiration from the human brain to imagine how massively parallel networks of simple neuron-like processors might learn and adapt from experience. There have been waves of excite- ment in this topic as new mathematical methods were developed. In 1958, Frank Rosenblatt proposed a learning device called the Perceptron, which had the amazing property that it could learn any task that one could program it by hand to perform. The enthusi- asm for this device vanished when Marvin Minsky and Seymour Papert performed a careful analysis showing limitations on what the device could be programmed to do, both in principle and in practice.