Adan: Adaptive Nesterov Momentum Algorithm for Faster Optimizing Deep Models
Top Cited Papers
- 4 July 2024
- journal article
- research article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 46 (12) , 9508-9520
Abstract
In deep learning, different kinds of deep networks typically need different optimizers, which have to be chosen after multiple trials, making the training process inefficient. To relieve this issue and consistently improve the model training speed across deep networks, we propose the ADAptive Nesterov momentum algorithm, Adan for short. Adan first reformulates the vanilla Nesterov acceleration to develop a new Nesterov momentum estimation (NME) method, which avoids the extra overhead of computing gradient at the extrapolation point. Then Adan adopts NME to estimate the gradient's first- and second-order moments in adaptive gradient algorithms for convergence acceleration. Besides, we prove that Adan finds an $\epsilon$ -approximate first-order stationary point within $\mathcal {O}(\epsilon ^{-3.5})$ stochastic gradient complexity on the non-convex stochastic problems (e.g., deep learning problems), matching the best-known lower bound. Extensive experimental results show that Adan consistently surpasses the corresponding SoTA optimizers on vision, language, and RL tasks and sets new SoTAs for many popular networks and frameworks, e.g., ResNet, ConvNext, ViT, Swin, MAE, DETR, GPT-2, Transformer-XL, and BERT. More surprisingly, Adan can use half of the training cost (epochs) of SoTA optimizers to achieve higher or comparable performance on ViT, GPT-2, MAE, etc , and also shows great tolerance to a large range of minibatch size, e.g., from 1 k to 32 k.
Keywords
All Related Versions
Funding Information
- National Key R&D Program of China (2022ZD0160300)
- NSF China (62276004)
- Qualcomm
- Singapore Ministry of Education (MOE) Academic Research Fund Tier 1
This publication has 36 references indexed in Scilit:
- SSCNet: learning-based subspace clusteringVisual Intelligence, 2024
- DeiT III: Revenge of the ViT2022
- Lower bounds for non-convex stochastic optimizationMathematical Programming, 2022
- Multi-way backpropagation for training compact deep neural networksNeural Networks, 2020
- Accelerated Optimization for Machine Learning2020
- Deep Networks with Stochastic Depth2016
- Microsoft COCO: Common Objects in Context2014
- Improvements to Deep Convolutional Neural Networks for LVCSR2013
- On-line Learning and Stochastic Approximations1999
- Some methods of speeding up the convergence of iteration methodsUSSR Computational Mathematics and Mathematical Physics, 1964