A new regret analysis for Adam-type algorithms
share
In this paper, we focus on a theory-practice gap for Adam and its variants (AMSgrad, AdamNC, etc.). In practice, these algorithms are used with a constant first-order moment parameter β_1 (typically between 0.9 and 0.99). In theory, regret guarantees for online convex optimization require a rapidly decaying β_1→0 schedule. We show that this is an artifact of the standard analysis and propose a novel framework that allows us to derive optimal, data-dependent regret bounds with a constant β_1, without further assumptions. We also demonstrate the flexibility of our analysis on a wide range of different algorithms and settings.
READ FULL TEXT