TL;DR: Knowing the next training checkpoint in advance (“lookahead”) helps to set the learning rate. In the limit, the classical square-root schedule appears on the horizon.

We show that learning-rate schedules for large model training behave surprisingly similar to a performance bound from non-smooth convex optimization theory. We provide a bound for the constant schedul...

Getting the correct bibtex entry for a conference paper (e.g. published at NeurIPS, ICML, ICLR) is annoyingly hard: if you search for the title, you will often find a link to arxiv or to the pdf file,...