Key points are not available for this paper at this time.
The popular backpropagation algorithm for training neural nets is a special case of an earlier principle of significance feedback, which in turn has much in common with Selfridge’s “Pandemonium” and a connection with McCulloch’s “redundancy of potential command”. Ways in which the effects might operate in real neural nets are reviewed, and the ideas are related to the current interest in heterogeneous agents. The tendency to restrict attention to numerical optimisation is regretted.
Alex M. Andrew (Sat,) studied this question.