It’s been a year, and I honestly dont remember reading the paper. I read through the TF implementation that was linked in the post, and I incorrectly assumed the code was for results in the paper. Had I spent a few more minutes, I would’ve know that it was done on MNIST, CIFAR10 and Penn Tree Bank. I apologize for being “that” person.
That being said, by MNIST, I meant “experiments done on small datasets”, as I’ve been burned multiple times on picking up an MNIST / CIFAR10 paper and it not working on large scale. And by “equivalent”, that was bad wording, I wanted people to get an intuition of the paper, similar to how you described in the README.