3 ms·

Sloppy Use of Machine Learning Is Causing a ‘Reproducibility Crisis’ in Science

• meltyness 4y ago
  ...just like regular-old statistics have been, for quite a while.
  • lumost 4y ago
    Black box ml model take the to a new level however. You can fit a model to anything, even if it lacks a causal relationship. If you setup the test set in such a way that it overlaps with the train set- then you'll get a positive result. You may think that you haven't overlapped the sets... But in practice most datasets have large numbers of duplicates once some irrelevant random variable is removed (eg time).
• anamax 4y ago
  Pretty much all of "social science" is not-reproducible. It's probably been that way forever, but it is now becoming obvious. Biology is also prone to "not reproducible". None of the above is due to machine learning.
• pyuser583 4y ago
  It’s hard enough making a ML model consistent with itself.