Revisiting the Ordering of Channel and Spatial Attention: A Comprehensive Study on Sequential and Parallel Designs

TL;DR

The paper tackles the lack of principled guidance for fusing Channel Attention and Spatial Attention in neural networks. It introduces a unified framework evaluating 18 topologies across four classes (sequential, parallel, multi-scale, residual) on two general-vision and nine MedMNIST datasets, revealing a data-scale–method–performance coupling law. Specifically, it finds that for $N<1k$ few-shot cases, channel–multi-scale spatial cascaded structures excel; for $1k le N le 50k$ medium-scale cases, parallel learnable fusion is superior; and for $N>50k$ large-scale cases, dynamic gated parallel structures perform best. The study also shows that Spatial→Channel order is more stable for fine-grained tasks, and residual connections mitigate gradient vanishing, culminating in scenario-based design guidelines and an open-source code release to facilitate practical, data-aware attention module design.

Abstract

Attention mechanisms have become a core component of deep learning models, with Channel Attention and Spatial Attention being the two most representative architectures. Current research on their fusion strategies primarily bifurcates into sequential and parallel paradigms, yet the selection process remains largely empirical, lacking systematic analysis and unified principles. We systematically compare channel-spatial attention combinations under a unified framework, building an evaluation suite of 18 topologies across four classes: sequential, parallel, multi-scale, and residual. Across two vision and nine medical datasets, we uncover a "data scale-method-performance" coupling law: (1) in few-shot tasks, the "Channel-Multi-scale Spatial" cascaded structure achieves optimal performance; (2) in medium-scale tasks, parallel learnable fusion architectures demonstrate superior results; (3) in large-scale tasks, parallel structures with dynamic gating yield the best performance. Additionally, experiments indicate that the "Spatial-Channel" order is more stable and effective for fine-grained classification, while residual connections mitigate vanishing gradient problems across varying data scales. We thus propose scenario-based guidelines for building future attention modules. Code is open-sourced at https://github.com/DWlzm.

Figures (8)

• Figure 1: Searching for the optimal attention mechanism among various combinations of channel and spatial attention.
• Figure 2: Structure Diagram of Basic Components
• Figure 3: Structure Diagram of Sequential Mode
• Figure 4: Structure Diagram of Parallel Mode
• Figure 5: Structure Diagram of Residual Connection Pattern