In recent years, the Vision Transformer (ViT) has garnered significant attention within the computer vision community. However, the core component of ViT, Self-Attention, lacks explicit spatial priors and suffers from quadratic computational complexity, limiting its applicability. To address these issues, we have proposed RMT, a robust vision backbone with explicit spatial priors for general purposes. RMT utilizes Manhattan distance decay to introduce spatial information and employs a horizontal and vertical decomposition attention method to model global information. Building on the strengths of RMT, Euclidean enhanced Vision Transformer (EVT) is an expanded version that incorporates several key improvements. Firstly, EVT uses a more reasonable Euclidean distance decay to enhance the modeling of spatial information, allowing for a more accurate representation of spatial relationships compared to the Manhattan distance used in RMT. Secondly, EVT abandons the decomposed attention mechanism featured in RMT and instead adopts a simpler spatially-independent grouping approach, providing the model with greater flexibility in controlling the number of tokens within each group. By addressing these modifications, EVT offers a more sophisticated and adaptable approach to incorporating spatial priors into the Self-Attention mechanism, thus overcoming some of the limitations associated with RMT and further enhancing its applicability in various computer vision tasks. Extensive experiments on Image Classification, Object Detection, Instance Segmentation, and Semantic Segmentation demonstrate that EVT exhibits exceptional performance. Without additional training data, EVT achieves 86.6% top1-acc on ImageNet-1 k.

The Softmax attention mechanism in Transformer models is notoriously computationally expensive due to its quadratic complexity, posing significant challenges in vision applications. In contrast, linear attention offers a far more efficient solution by reducing the complexity to linear levels. However, linear attention often suffers significant performance degradation compared to Softmax attention. Our experiments indicate that this performance drop stems from the low-rank nature of linear attention’s output feature map, which hinders its ability to adequately model complex spatial information. To address this low-rank dilemma, we conduct rank analysis from two perspectives: the KV buffer and the output features. Consequently, we introduce Rank-Augmented Linear Attention (RALA), which rivals the performance of Softmax attention while maintaining linear complexity and high efficiency. Building upon RALA, we construct the Rank-Augmented Vision Linear Transformer (RAVLT). Extensive experiments demonstrate that RAVLT achieves excellent performance across various vision tasks. Specifically, without using any additional labels, data, or supervision during training, RAVLT achieves an 84.4% Top-1 accuracy on ImageNet-1k with only 26M parameters and 4.6G FLOPs. This result significantly surpasses previous linear attention mechanisms, fully illustrating the potential of RALA. Code will be available at https://github.com/qhfan/RALA.

As the core operator of Transformers, Softmax Attention exhibits excellent global modeling capabilities. However, its quadratic complexity limits its applicability to vision tasks. In contrast, Linear Attention shares a similar formulation with Softmax Attention while achieving linear complexity, enabling efficient global information modeling. Nevertheless, Linear Attention suffers from a significant performance degradation compared to standard Softmax Attention. In this paper, we analyze the underlying causes of this issue based on the formulation of Linear Attention. We find that, unlike Softmax Attention, Linear Attention entirely disregards the magnitude information of the Query. This prevents the attention score distribution from dynamically adapting as the Query scales. As a result, despite its structural similarity to Softmax Attention, Linear Attention exhibits a significantly different attention score distribution. Based on this observation, we propose Magnitude-Aware Linear Attention (MALA), which modifies the computation of Linear Attention to fully incorporate the Query's magnitude. This adjustment allows MALA to generate an attention score distribution that closely resembles Softmax Attention while exhibiting a more well-balanced structure. We evaluate the effectiveness of MALA on multiple tasks, including image classification, object detection, instance segmentation, semantic segmentation, natural language processing, speech recognition, and image generation. Our MALA achieves strong results on all of these tasks. Code will be available at https://github.com/qhfan/MALA

In recent years, Transformers have achieved remarkable progress in computer vision tasks. However, their global modeling often comes with substantial computational overhead, in stark contrast to the human eye's efficient information processing. Inspired by the human eye's sparse scanning mechanism, we propose a Sparse Scan Self-Attention mechanism (S3A). This mechanism predefines a series of Anchors of Interest for each token and employs local attention to efficiently model the spatial information around these anchors, avoiding redundant global modeling and excessive focus on local information. This approach mirrors the human eye's functionality and significantly reduces the computational load of vision models. Building on S3A, we introduce the Sparse Scan Vision Transformer (SSViT). Extensive experiments demonstrate the outstanding performance of SSViT across a variety of tasks. Specifically, on ImageNet classification, without additional supervision or training data, SSViT achieves top-1 accuracies of 84.4%/85.7% with 4.4G/18.2G FLOPs. SSViT also excels in downstream tasks such as object detection, instance segmentation, and semantic segmentation. Its robustness is further validated across diverse datasets.

Diffusion Transformer (DiT), a promising diffusion model for visual generation, demonstrates impressive performance but incurs significant computational overhead. Intriguingly, analysis of pre-trained DiT models reveals that global self-attention is often redundant, predominantly capturing local patterns—highlighting the potential for more efficient alternatives. In this paper, we revisit convolution as an alternative building block for constructing efficient and expressive diffusion models. However, naively replacing self-attention with convolution typically results in degraded performance. Our investigations attribute this performance gap to the higher channel redundancy in ConvNets compared to Transformers. To resolve this, we introduce a compact channel attention mechanism that promotes the activation of more diverse channels, thereby enhancing feature diversity. This leads to Diffusion ConvNet (DiCo), a family of diffusion models built entirely from standard ConvNet modules, offering strong generative performance with significant efficiency gains. On class-conditional ImageNet generation benchmarks, DiCo-XL achieves an FID of 2.05 at 256$\times$256 resolution and 2.53 at 512$\times$512, with a **2.7$\times$** and **3.1$\times$** speedup over DiT-XL/2, respectively. Furthermore, experimental results on MS-COCO demonstrate that the purely convolutional DiCo exhibits strong potential for text-to-image generation.

Tingkai Liu, Yunzhe Tao, Haogeng Liu, Qihang Fang, Ding Zhou, Huaibo Huang, Ran He, Hongxia Yang. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Vision Transformer (ViT) has gained increasing attention in the computer vision community in recent years. How-ever, the core component of ViT, Self-Attention, lacks ex-plicit spatial priors and bears a quadratic computational complexity, thereby constraining the applicability of ViT. To alleviate these issues, we draw inspiration from the re-cent Retentive Network (RetNet) in the field of NLP, and propose RMT, a strong vision backbone with explicit spa-tial prior for general purposes. Specifically, we extend the RetNet's temporal decay mechanism to the spatial do-main, and propose a spatial decay matrix based on the Manhattan distance to introduce the explicit spatial prior to Self-Attention. Additionally, an attention decomposition form that adeptly adapts to explicit spatial prior is proposed, aiming to reduce the computational burden of modeling global information without disrupting the spa-tial decay matrix. Based on the spatial decay matrix and the attention decomposition form, we can flexibly integrate explicit spatial prior into the vision backbone with lin-ear complexity. Extensive experiments demonstrate that RMT exhibits exceptional performance across various vision tasks. Specifically, without extra training data, RMT achieves 84.8% and 86.1% top-l acc on ImageNet-lk with 27MI4.5GFLOPs and 96M/18.2GFLOPs. For downstream tasks, RMT achieves 54.5 box AP and 47.2 mask AP on the COCO detection task, and 52.8 mloU on the ADE20K se-mantic segmentation task.

Recent advancements in vision backbones have significantly improved their performance by simultaneously modeling images’ local and global contexts. However, the bidirectional interaction between these two contexts has not been well explored and exploited, which is important in the human visual system. This paper proposes a **F**ully **A**daptive **S**elf-**A**ttention (FASA) mechanism for vision transformer to model the local and global information as well as the bidirectional interaction between them in context-aware ways. Specifically, FASA employs self-modulated convolutions to adaptively extract local representation while utilizing self-attention in down-sampled space to extract global representation. Subsequently, it conducts a bidirectional adaptation process between local and global representation to model their interaction. In addition, we introduce a fine-grained downsampling strategy to enhance the down-sampled self-attention mechanism for finer-grained global perception capability. Based on FASA, we develop a family of lightweight vision backbones, **F**ully **A**daptive **T**ransformer (FAT) family. Extensive experiments on multiple vision tasks demonstrate that FAT achieves impressive performance. Notably, FAT accomplishes a **77.6%** accuracy on ImageNet-1K using only **4.5M** parameters and **0.7G** FLOPs, which surpasses the most advanced ConvNets and Transformers with similar model size and computational costs. Moreover, our model exhibits faster speed on modern GPU compared to other models.