Mamba-papers

日期: September 6th 2024, 5:24:25 am
期刊: xxx

Mamba: Linear-Time Sequence Modeling with Selective State Spaces

Abstract

基础模型目前为深度学习领域大多数令人兴奋的应用提供了动力,这些模型几乎都基于 Transformer 架构及其核心注意力模块。为了解决 Transformer 在长序列上的计算效率低下问题,人们开发了许多亚二次时间架构,如线性注意力、门控卷积和递归模型以及结构化状态空间模型(SSM),但它们在语言等重要模态上的表现不如注意力。我们发现,此类模型的一个关键弱点是无法进行基于内容的推理,因此做出了几项改进。首先,只需让 SSM 参数成为输入的函数,就能解决它们在离散模态方面的弱点,使模型能够根据当前标记,有选择地沿序列长度维度传播或遗忘信息。其次,尽管这种变化阻碍了高效卷积的使用,我们还是设计了一种硬件感知的并行递归模式算法。我们将这些选择性 SSM 集成到一个简化的端到端神经网络架构中,该架构没有注意力,甚至没有 MLP 块(Mamba)。Mamba 具有快速推理(吞吐量比 Transformers 高 5 倍)和序列长度线性伸缩的特点,其性能在实际数据中可提高到百万长度序列。作为通用序列模型的骨干,Mamba 在语言、音频和基因组学等多种模式中都达到了最先进的性能。在语言建模方面,无论是预训练还是下游评估,我们的 Mamba-3B 模型都优于同等规模的 Transformers,并能与两倍于其规模的 Transformers 相媲美。

结构

Mamba1

U-shaped Vision Mamba for Single Image Dehazing

![U-shaped Vision Mamba for Single Image Dehazing1](../images/Mamba-papers/U-shaped Vision Mamba for Single Image Dehazing1.png)

VMamba: Visual State Space Model

VMamba1

Mamba2

U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmentation

![U-Mamba Enhancing Long-range Dependency for Biomedical Image Segmentation1](../images/Mamba-papers/U-Mamba Enhancing Long-range Dependency for Biomedical Image Segmentation1.png)

Swin-UMamba: Mamba-based UNet with ImageNet-based pretraining

![Swin-UMamba Mamba-based UNet with ImageNet-based pretraining](../images/Mamba-papers/Swin-UMamba Mamba-based UNet with ImageNet-based pretraining.png)

Semi-Mamba-UNet: Pixel-Level Contrastive Cross-Supervised Visual Mamba-based UNet for Semi-Supervised Medical Image Segmentation

![Semi-Mamba-UNet Pixel-Level Contrastive Cross-Supervised Visual Mamba-based UNet for Semi-Supervised Medical Image Segmentation1](../images/Mamba-papers/Semi-Mamba-UNet Pixel-Level Contrastive Cross-Supervised Visual Mamba-based UNet for Semi-Supervised Medical Image Segmentation1.png)

SegMamba: Long-range Sequential Modeling Mamba For 3D Medical Image Segmentation

![SegMamba Long-range Sequential Modeling Mamba For 3D Medical Image Segmentation1](../images/Mamba-papers/SegMamba Long-range Sequential Modeling Mamba For 3D Medical Image Segmentation1.png)

![SegMamba Long-range Sequential Modeling Mamba For 3D Medical Image Segmentation2](../images/Mamba-papers/SegMamba Long-range Sequential Modeling Mamba For 3D Medical Image Segmentation2.png)

nnMamba: 3D Biomedical Image Segmentation, Classification and Landmark Detection with State Space Model

![nnMamba 3D Biomedical Image Segmentation, Classification and Landmark Detection with State Space Model1](../images/Mamba-papers/nnMamba 3D Biomedical Image Segmentation, Classification and Landmark Detection with State Space Model1.png)

Mamba-UNet: UNet-Like Pure Visual Mamba for Medical Image Segmentation

![Mamba-UNet UNet-Like Pure Visual Mamba for Medical Image Segmentation1](../images/Mamba-papers/Mamba-UNet UNet-Like Pure Visual Mamba for Medical Image Segmentation1.png)

![Mamba-UNet UNet-Like Pure Visual Mamba for Medical Image Segmentation2](../images/Mamba-papers/Mamba-UNet UNet-Like Pure Visual Mamba for Medical Image Segmentation2.png)

Weak-Mamba-UNet: Visual Mamba Makes CNN and ViT Work Better for Scribble-based Medical Image Segmentation

![Weak-Mamba-UNet Visual Mamba Makes CNN and ViT Work Better for Scribble-based Medical Image Segmentation1](../images/Mamba-papers/Weak-Mamba-UNet Visual Mamba Makes CNN and ViT Work Better for Scribble-based Medical Image Segmentation1.png)

VM-UNet: Vision Mamba UNet for Medical Image Segmentation

![VM-UNet Vision Mamba UNet for Medical Image Segmentation1](../images/Mamba-papers/VM-UNet Vision Mamba UNet for Medical Image Segmentation1.png)