Adaptivity without Compromise: A Momentumized, Adaptive, Dual Averaged Gradient Method for Stochastic Optimization
This addresses the challenge of adaptive methods performing poorly in certain deep learning optimization problems, though it appears incremental as it builds on existing AdaGrad family methods.
The paper tackled the problem of improving adaptive gradient methods for stochastic optimization by introducing MADGRAD, which matches or outperforms SGD and ADAM in test set performance across vision and NLP tasks.
We introduce MADGRAD, a novel optimization method in the family of AdaGrad adaptive gradient methods. MADGRAD shows excellent performance on deep learning optimization problems from multiple fields, including classification and image-to-image tasks in vision, and recurrent and bidirectionally-masked models in natural language processing. For each of these tasks, MADGRAD matches or outperforms both SGD and ADAM in test set performance, even on problems for which adaptive methods normally perform poorly.