TIL: People are able to predict studies from people in their fields who probably a have a similar bias.
I was surprised to read about the low reproducibility rate, how can you call it social "science" if you can't reproduce it?
I was surprised to read about the low reproducibility rate, how can you call it social "science" if you can't reproduce it?
NB: this is not specific to the social sciences. psychology, cognitive sciences, medicine(!), and other fields have a 'replication crisis' also. there are some really good papers on this. i am not aware of systematic queries into CS/ML/DL papers, but from my own experience: when you pick up a paper that claims 9x% accuracy, the chance that you are able to replicate that number is in line with replication % in other fields.
From all my friends who pursued STEM, I assume this is happening because of the pressure on people pursuing higher education at all costs, not finding enough qualified jobs, getting stuck in academia and printing low quality papers to get a promotion / keep getting another public grant.
If you add in to the mix results influenced by the current political climate, what your peers think, what your sponsors want to prove for financial gain, the situation gets bleak very fast.
Irreproducible papers are the abandoned OSS projects of software engineers in private tech + corruption + public money, a recipe for disaster.
OSS is often someone learning, or who has an itch to scratch, and put some code out there that may or may not solve your problem or work right. The "lack of warranty" clause in most open-source licensing is really important here, because it's largely designed to say "hey, I did a thing but make no guarantees so use at your own risk", whereas a published paper says "hey, a number of experts and people who should know all agree that this is a thoroughly-researched and well-thought-out position, and you can probably consider it to be true (or nearly so) and base some of your decisions on it." I think that change in context is really important to consider when making the analogy here.
This is a myth. Modern "peer review" simply means checking to see if the claim is interesting and if the proper Word or LaTeX template was used. Peer review meant replication in the distant past before the modern Grants and Impact Factor system was built.
I got peer review from two reviewers on a paper submitted to Frontiers last week--it consisted of much more than that. Sure, the reviewers couldn't replicate our exact work themselves, but the paper's methodology was also heavily considered, and both reviewers asked for modifications to be made on different parts of the paper.
It's an imperfect system, but it's not as if there's a rubber stamp floating around. The goals of the the peer review system vs. open source are fundamentally different, and the way the products for both are treated should reflect that.