Schedules & Schedule-Free Learning (Aaron Defazio, 12.06.2024 at UCLA)

Ernest Ryu · 64:04

In this talk, Aaron Defazio argues that the textbook gap between learning-rate theory (use a constant step, average the iterates) and real practice (decaying schedules, return the last iterate) goes away once you see...

Read the full summary on tuber

Redirecting...