Couple of days ago I was intrigued by a paper with a very loud title: “Dropout is a special case of the stochastic delta rule: faster and more accurate deep learning” (link: https://arxiv.org/abs/1808.03578).
The paper has received a bit of attention, even Jeremy twitted about it.
But one guy has replied to him:
Bas Veeling (https://twitter.com/basveeling –– PhD from Amsterdam) created an issue on github:
Two hours later authors of the paper have replied:
And Jeremy has twitted about the paper again.
So here is a moral: don’t trust papers without the code.