[WnZ39p0Dgydaz1]

> Of late he feels like all the activity of himself and his peers is just playing the Science Game: varying some variable with infinite degrees of freedom and then throwing statistics at it until you get that reportable p-value and write up a narrative short story around it.

Just like ML research, we just skip the p-values.

[belval]

The classic "we achieve SOTA on ImageNet by using a novel training procedure" where they just played with learning rate schedules until they got 0.1% over previous SOTA.

To be fair to there are a lot of smells in DL papers, usually you can tell whether an approach is worth your time by looking at code availability, lab, previous publications and the conference where it was published.

[robwwilliams]

Good zinger.