News for hackers. Stuff that matters.
Re:An overview of the theory of overparameterized machine le... by Salgat on Tuesday September 21, 2021 @12:14PM (#28606555)
To add to this, there's a misleading phenomenon that first occurs where the performance actually gets worse with too much data/parameters/epochs, but oddly improves again if you throw even more at the model.
Re:An overview of the theory of overparameterized machine le... by jointpdf on Tuesday September 21, 2021 @01:00PM (#28607094)
For the interested, this phenomenon is known as (deep) double descent:

https://openai.com/blog/deep-double-descent/

https://www.lesswrong.com/posts/FRv7ryoqtvSuqBxuT/understand...

(Edit: Oh, the definition appears in the abstract of the linked paper.)

Re:An overview of the theory of overparameterized machine le... by nexuist on Tuesday September 21, 2021 @12:31PM (#28606752)
Is this the ML equivalent of Dunning–Kruger effect? A model with a bit of data is too afraid of being wrong to be overconfident. A model with a bit more data is overconfident in itself and gets things wrong. Finally, a model with tons and tons of data understands the complexity of the problem set and once again becomes too afraid of being wrong.
Re:An overview of the theory of overparameterized machine le... by visarga on Tuesday September 21, 2021 @12:55PM (#28607028)
Model confidence as reported by softmax probability scores is notoriously noisy and miscalibrated. With larger models and more data the confidence estimation gets more nuanced.