Research map / Architectures and efficiency / Attention and transformers
Stripe Observation Guided Inference Cost-free Attention Mechanism
Zhongzhan Huang, Shanshan Zhong, Wushao Wen, Jinghui Qin, Liang Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ASR (SE/ECA/SRM) re-parameterized attention to ResNet at no inference cost | ResNet34/50/101 · ImageNet-1k val | top-1 accuracy: 76.13 -> 76.87 | Improved |
| add ASR to ShuffleNetV2 with SE | ShuffleNetV2 · CIFAR10 | top-1 accuracy: 91.68 -> 91.68 | No clear effect |
| add ASR on top of an existing attention module | SE in ResNet164 · CIFAR100 | top-1 accuracy: 75.29 -> 77.32 | Improved |
| replace standalone attention with ASR counterpart | SRM in ResNet164 · CIFAR100 | top-1 accuracy: 75.38 -> 75.23 | Hurt |
| add ASR to RepVGG and ACNet | RepVGG, ACNet · STL10 / CIFAR100 | top-1 accuracy: 83.54 -> 86.13 | Improved |
| insert ASR multiple times (delta from 1 to 2, then 4) (ablation) | ResNet164 ASR (SE) · CIFAR100 | top-1 accuracy: 75.36 -> 75.87 -> 75.22 | Mixed |
| make input psi learnable rather than fixed constant or random (ablation) | ASR (SE) · CIFAR100 | top-1 accuracy: 75.18 -> 75.36 | Improved |
| add ASR under constant noise injected into BN layers | ResNet164 · CIFAR100 noise (0.5,0.5) | top-1 accuracy: 35.77 -> 72.18 | Improved |
| use ASR with constant-vector attention in object detection and segmentation | Mask R-CNN ResNet50 · COCO2017 segmentation AP | AP: 35.2 -> 35.1 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- ShuffleNetV2 — ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design (ECCV 2018) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Understanding the Differences in Foundation Models: Attention, State Space Models, and Recurrent Neural Networks (NeurIPS 2024 · similarity 0.83)
- BA-SAM: Scalable Bias-Mode Attention Mask for Segment Anything Model (CVPR 2024 · similarity 0.82)
- SpargeAttention: Accurate and Training-free Sparse Attention Accelerating Any Model Inference (ICML 2025 · similarity 0.82)
- Self-Attention-Based Contextual Modulation Improves Neural System Identification (ICLR 2025 · similarity 0.82)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (NeurIPS 2025 · similarity 0.82)
- SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
