The more relevant link is https://keplerlounge.com/neural-computation/2020/01/16/compl.... Where the author discusses why this repos exists. This isn't a very good way to do AutoDiff and isn't exact(Look at standard reverse mode AutoDiff or Julia or Swift's implementations instead). The point of the repo is that the author believes it is a biologically plausible mechanism for neurons to be trained. Whether it is or isn't is the interesting question.