An approximation of gradients used to speed up training in deep learning by decoupling layers.
Lex Fridman