Distance-based, parameter-free step-size schedules for AdaGrad and Adam give provable convex O(1/sqrt(T)) rates and competitive deep-learning results, with caveats about initialization and theory.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Towards Simple and Provable Parameter-Free Adaptive Gradient Methods
Distance-based, parameter-free step-size schedules for AdaGrad and Adam give provable convex O(1/sqrt(T)) rates and competitive deep-learning results, with caveats about initialization and theory.