Comment by fxtentacle

These techniques are not new. And the reason why they’re usually not used is on page 9 in the paper. They require about 10x as many training iterations.