3 ms·

It’s been a while for me, but I’m pretty sure there are techniques for approximating gradients for step functions, although the paper may not have mentioned the

• fxtentacle 7y ago
  In the case I evaluated, the authors later admitted to using supervised training for initializing their supposedly unsupervised learning methods. Also, I was reviewing their public source code release and there was no approximation. That part of their loss function had simply never worked, but they had not noticed prior to the release of their paper. And due to them training slightly different from what they described in the paper, the AI worked competitively well nonetheless.