Adan: Adaptive Nesterov Momentum Algorithm for Faster Optimizing Deep Models

By Xingyu Xie, Pan Zhou, Huan Li, Zhouchen Lin, Shuicheng Yan. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022

313 citations recorded by the source corpus.

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 <inline-formula><tex-math notation="LaTeX">$\epsilon$</tex-math><alternatives><mml:math><mml:mi>ε</mml:mi></mml:math><inline-graphic xlink:href="lin-ieq1-3423382.gif"/></alternatives></inline-formula>-approximate first-order stationary point within <inline-formula><tex-math notation="LaTeX">$\mathcal {O}(\epsilon ^{-3.5})$</tex-math><alternatives><mml:math><mml:mrow><mml:mi mathvariant="script">O</mml:mi><mml:mo>(</mml:mo><mml:msup><mml:mi>ε</mml:mi><mml:mrow><mml:mo>-</mml:mo><mml:mn>3</mml:mn><mml:mo>.</mml:mo><mml:mn>5</mml:mn></mml:mrow></mml:msup><mml:mo>)</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="lin-ieq2-3423382.gif"/></alternatives></inline-formula> 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, <italic>etc</italic>, and also shows great tolerance to a large range of minibatch size, e.g., from 1 k to 32 k.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/81521a80f7aa939281863c15b42e446ff5a0e65a

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Sora: A Review on Background, Technology, Limitations, and Opportunities of Large Vision Models (2024, methodology): “Furthermore, they also leverage a bag of training strategies like the Adan optimizer [57], Min-SNR weighting [58], dynamic masking ratio, and an improved power-cosine weighting for classifier-free guidance [59].”
A modified Adam algorithm for deep neural network optimization (2023, methodology): “ResNet, ConvNext (2022) [52] ImageNet ADAN (Top-1 accuracy on ViT Small = 80.”
Masked Diffusion Transformer is a Strong Image Synthesizer (2023, methodology): “Fre’chet Inception Distance (FID) [20], sFID [33], Inception Score (IS) [40], Precision and Recall [29].”