851300ee84c2b80ed40f51ed26d866fc-AuthorFeedback.pdf

For the backward pass in the standard GIM model, we block the flow of27 gradients to the previous module. We can express this using the gradient blocking operator as defined in the draft28 (GradientBlock(x), x, GradientBlock(x), 0),such thatct = gar(GradientBlock(zt),ht 1).