Compare/contrast various local rules that approximate gradient descent on a global learning cost function and how they are implemented in hardware. Eg electrical circuits that implement "coupled learning" to do supervised learning tasks on their own, without a processor.  Seeking speakers!