Deep Incubation: Training Large Models by Divide-and-Conquering
Zanlin Ni, Yulin Wang, Jiangwei Yu, Haojun Jiang, Yue Cao, Gao Huang
摘要
Recent years have witnessed a remarkable success of large deep learning models. However, training these models is challenging due to high computational costs, painfully slow convergence, and overfitting issues. In this paper, we present Deep Incubation, a novel approach that enables the efficient and effective training of large models by dividing them into smaller sub-modules which can be trained separately and assembled seamlessly. A key challenge for implementing this idea is to ensure the compatibility of the independently trained sub-modules. To address this issue, we first introduce a global, shared meta model, which is leveraged to implicitly link all the modules together, and can be designed as an extremely small network with negligible computational overhead. Then we propose a module incubation algorithm, which trains each sub-module to replace the corresponding component of the meta model and accomplish a given learning task. Despite the simplicity, our approach effectively encourages each sub-module to be aware of its role in the target large model, such that the finally-learned sub-modules can collaborate with each other smoothly after being assembled. Empirically, our method outperforms end-to-end (E2E) training in terms of both final accuracy and training efficiency. For example, on top of ViT-Huge, it improves the accuracy by 2.7% on Ima-geNet or achieves similar performance with 4× less training time. Notably, the gains are significant for downstream tasks as well (e.g., object detection and image segmentation on COCO and ADE20K). Code is available at https:// github.com/LeapLabTHU/Deep-Incubation .
问 Lune
几分钟读透这篇 Paper,不用再花一下午。
Lune 能读完全文,也能读引用它的顶会 Paper,讲清它的结论、证据和局限,每一点都注明出自哪篇 Paper。不用硬啃,也不靠猜。
已有 7 篇顶会 Paper 引用了它。别等审稿人问起,先弄清它们发现了什么。
脉络
- CoAtNet: Marrying Convolution and Attention for All Data Sizes 引用了 An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- CoAtNet: Marrying Convolution and Attention for All Data Sizes 引用了 Language Models are Few-Shot Learners
- FLatten Transformer: Vision Transformer using Focused Linear Attention 引用了 EfficientTrain: Exploring Generalized Curriculum Learning for Training Visual Backbones
- FLatten Transformer: Vision Transformer using Focused Linear Attention 引用了 An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- EfficientTrain: Exploring Generalized Curriculum Learning for Training Visual Backbones 引用了 An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- EfficientTrain: Exploring Generalized Curriculum Learning for Training Visual Backbones 引用了 A ConvNet for the 2020s
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale 引用了 A Simple Framework for Contrastive Learning of Visual Representations
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale 引用了 Language Models are Few-Shot Learners
- A ConvNet for the 2020s 引用了 CoAtNet: Marrying Convolution and Attention for All Data Sizes
- A ConvNet for the 2020s 引用了 An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- Demystify Mamba in Vision: A Linear Attention Perspective 引用了 FLatten Transformer: Vision Transformer using Focused Linear Attention
- Demystify Mamba in Vision: A Linear Attention Perspective 引用了 An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- Demystify Mamba in Vision: A Linear Attention Perspective 引用了 A ConvNet for the 2020s
- Demystify Mamba in Vision: A Linear Attention Perspective 引用了 ENAT: Rethinking Spatial-temporal Interactions in Token-based Image Synthesis
- ENAT: Rethinking Spatial-temporal Interactions in Token-based Image Synthesis 引用了 Demystify Mamba in Vision: A Linear Attention Perspective
- ENAT: Rethinking Spatial-temporal Interactions in Token-based Image Synthesis 引用了 FLatten Transformer: Vision Transformer using Focused Linear Attention
- ENAT: Rethinking Spatial-temporal Interactions in Token-based Image Synthesis 引用了 Language Models are Few-Shot Learners
- GSVA: Generalized Segmentation via Multimodal Large Language Models 引用了 EfficientTrain: Exploring Generalized Curriculum Learning for Training Visual Backbones
- GSVA: Generalized Segmentation via Multimodal Large Language Models 引用了 An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- GSVA: Generalized Segmentation via Multimodal Large Language Models 引用了 Language Models are Few-Shot Learners
引用它的顶会 Paper7
- FLatten Transformer: Vision Transformer using Focused Linear AttentionDongchen Han, Xuran Pan, Yizeng Han, Shiji Song 等ICCV 2023 · 358 次引用
- Demystify Mamba in Vision: A Linear Attention PerspectiveDongchen Han, Ziyi Wang, Zhuofan Xia, Yizeng Han 等NeurIPS 2024 · 287 次引用
- GSVA: Generalized Segmentation via Multimodal Large Language ModelsZhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan 等CVPR 2024 · 42 次引用
- EfficientTrain: Exploring Generalized Curriculum Learning for Training Visual BackbonesYulin Wang, Yang Yue, Rui Lu, Tianjiao Liu 等ICCV 2023 · 39 次引用
- ENAT: Rethinking Spatial-temporal Interactions in Token-based Image SynthesisZanlin Ni, Yulin Wang, Renping Zhou, Yizeng Han 等NeurIPS 2024 · 15 次引用
它引用的 Paper18
- Language Models are Few-Shot LearnersTom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah 等NeurIPS 2020 · 64,255 次引用
- A Simple Framework for Contrastive Learning of Visual RepresentationsTing Chen, Simon Kornblith, Mohammad Norouzi, Geoffrey E. HintonICML 2020 · 24,064 次引用
- An Image is Worth 16x16 Words: Transformers for Image Recognition at ScaleAlexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn 等ICLR 2021 · 21,477 次引用
- A ConvNet for the 2020sZhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer 等CVPR 2022 · 6,782 次引用
- CoAtNet: Marrying Convolution and Attention for All Data SizesZihang Dai, Hanxiao Liu, Quoc V. Le, Mingxing TanNeurIPS 2021 · 1,747 次引用
相关 Paper
- Masked Autoencoders Are Scalable Vision LearnersKaiming He, Xinlei Chen, Saining Xie, Yanghao Li 等CVPR 2022
- Memory Efficient Meta-Learning with Large ImagesJohn Bronskill, Daniela Massiceti, Massimiliano Patacchiola, Katja Hofmann 等NeurIPS 2021 · 21 次引用
- Masked Image Residual Learning for Scaling Deeper Vision TransformersGuoxi Huang, Hongtao Fu, Adrian G. BorsNeurIPS 2023 · 10 次引用
- A Multi-Level Framework for Accelerating Training Transformer ModelsLongwei Zou, Han Zhang, Yangdong DengICLR 2024 · 3 次引用
- MASS: MoErging through Adaptive Subspace SelectionDonato Crisostomi, Alessandro Zirilli, Antonio Andrea Gargiulo, Maria Sofia Bucarelli 等ICLR 2026 · 4 次引用
