Reading the openreview link, the current understanding is that this approach is dramatically more computationally intensive than standard backprop - limiting it

Reading the openreview link, the current understanding is that this approach is dramatically more computationally intensive than standard backprop - limiting its utility.