Adaptivity without Compromise: A Momentumized, Adaptive, Dual Averaged Gradient Method for Stochastic Optimization
arXiv:2101.11075
Abstract
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.
References in corpus (9)
- Adam: A Method for Stochastic Optimization
- On the Convergence of Adam and Beyond
- fastMRI: An Open Dataset and Benchmarks for Accelerated MRI
- AdaGrad stepsizes: Sharp convergence over nonconvex landscapes
- ZeRO: Memory Optimizations Toward Training Trillion Parameter Models
- Descending through a Crowded Valley - Benchmarking Deep Learning Optimizers
- A Simple Convergence Proof of Adam and Adagrad
- Online Adaptive Methods, Universality and Acceleration
- Offset Sampling Improves Deep Learning based Accelerated MRI Reconstructions by Exploiting Symmetry
Cited by in corpus (7)
- Descending through a Crowded Valley - Benchmarking Deep Learning Optimizers
- The Multi-Agent Behavior Dataset: Mouse Dyadic Social Interactions
- Sequential Training of Neural Networks with Gradient Boosting
- Noise-aware Physics-informed Machine Learning for Robust PDE Discovery
- Deep Symbolic Learning: Discovering Symbols and Rules from Perceptions
- Incremental Without Replacement Sampling in Nonconvex Optimization
- Leveraging Cross-Utterance Context For ASR Decoding