7 ms·
Speaking from my own areas of experience, I've been struck by how much some celebrated methods and results in machine learning, broadly speaking, basically amou
by kem 10y ago
Speaking from my own areas of experience, I've been struck by how much some celebrated methods and results in machine learning, broadly speaking, basically amount to tinkering without any hard generalizable proofs. In at least one case I'm aware of, there was a statistical proof later that made it clear the original researchers were close, but if they had a different dataset they would have come up with a different result. In fact, as I write this, I think I recall that different researchers did come up with different results, and it was in large part due to tinkering on different datasets.
So yes, even in computer science you run into similar things. Maybe not the same, but similar processes leading to similar problems.
- throwaway729 10y agoThe problem is that universities have a strong incentive to optimize for juicy press releases. A theorem -- unless it's 100 years old or otherwise somehow easy to relate to the general public -- does not have this effect. Winning a competition or "outperforming humans" does. Aside from bumping up the quality of K-12 science education by at least an order of magnitude or two, I'm not sure there's a good solution to this problem. It's the human condition to be impressed by some things and not others, regardless of their relative actual importance.
- thom_kaar 10y agoI am very interested in the paper with the mathematical proof. Would you link it or just mention its title?