Systems that emulate biological neural networks offer an efficient way of running AI algorithms, but they can’t be trained using the conventional approach. The symmetry of these ‘physical’ networks ...