Re:Implement your own source transformation AD with Julia
by throwgeorge on Thursday June 11, 2020 @12:48PM
(#23490731)
the gradient doesn't exist but subgradients do exist (at points of non-differentiability) and that is still useful (case in point as someone else mentions ReLU)
https://see.stanford.edu/materials/lsocoee364b/01-subgradien...