Modern science is rapidly approaching a crisis provoked by the widespread use of machine learning technologies. This statement was made at the American Association for the Advancement of Science conference in Washington, D. C. Genevera Allen, a statistician at Rice University.
Allen spoke about the serious problem associated with the so-called reproducibility crisis. Applying algorithms close to AI and poorly understanding the principles of their work, modern scientists often pay too much attention to "noise", which cannot be reproduced with repeated experiments.
“Researchers already have an understanding of the reproducibility crisis. I believe the root cause of the problem is the use of machine learning algorithms,”said Allen.
It often happens that the results of research carried out using machine learning look quite plausible, Allen said, however, as soon as research conducted with a large set of data appears, the old immediately begins to look inaccurate.
“The key problem with machine learning is that it finds patterns even where there are none at all. The only way out of this situation is to develop new algorithms capable of generating truly reliable and reproducible predictions,”says the statistician.
Kolesnikov Andrey