Hacker News new | ask | show | jobs
by esafak 303 days ago
Bayesian models solve this problem but they occupy model capacity which practitioners have traditionally preferred to devote to improving point estimates.
1 comments

I've always found this perspective remarkably misguided. Prediction performance is not everything; it can be extraordinarily powerful to have uncertainty estimates as well.
A lot of machine learning practitioners don't have the statistical sophistication to appreciate this, and as ML becomes increasingly "democratized" I expect the situation will worsen.