SEMA: a Scalable and Efficient Mamba like Attention via Token Localization and Averaging

2025-06-10

AI Summary

Traditional vision attention mechanisms suffer from two fundamental limitations: quadratic computational complexity and poor focus capability in linear variants. This paper proposes SEMA, a novel attention framework that first formally defines generalized attention and theoretically proves its inherent dispersion property. Leveraging this insight, SEMA introduces a dual-path “localization + averaging” architecture: token localization suppresses weight dispersion, while theoretically consistent arithmetic mean aggregation restores focusing ability—preserving linear complexity. Furthermore, SEMA integrates Mamba-inspired state-space modeling to enhance long-range dependency capture. On ImageNet-1K, SEMA significantly outperforms vision Mamba models of comparable parameter count; notably, its performance gain increases with larger image resolutions, demonstrating superior scalability and effectiveness.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Attention is the critical component of a transformer. Yet the quadratic computational complexity of vanilla full attention in the input size and the inability of its linear attention variant to focus have been challenges for computer vision tasks. We provide a mathematical definition of generalized attention and formulate both vanilla softmax attention and linear attention within the general framework. We prove that generalized attention disperses, that is, as the number of keys tends to infinity, the query assigns equal weights to all keys. Motivated by the dispersion property and recent development of Mamba form of attention, we design Scalable and Efficient Mamba like Attention (SEMA) which utilizes token localization to avoid dispersion and maintain focusing, complemented by theoretically consistent arithmetic averaging to capture global aspect of attention. We support our approach on Imagenet-1k where classification results show that SEMA is a scalable and effective alternative beyond linear attention, outperforming recent vision Mamba models on increasingly larger scales of images at similar model parameter sizes.

Problem

• Addresses quadratic complexity of vanilla full attention
• Solves linear attention's inability to focus effectively
• Proposes SEMA for scalable efficient attention in vision tasks

Innovation

• Token localization prevents attention dispersion
• Arithmetic averaging captures global attention aspects
• Scalable Mamba-like attention for vision tasks