Research map / Architectures and efficiency / Mixture of experts
ShiftAddViT: Mixture of Multiplication Primitives Towards Efficient Vision Transformer
Haoran You, Huihong Shi, Yipin Guo, Yingyan Celine 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 |
|---|---|---|---|
| Replace dense attention MatMuls with additions and use shifts/MoE for linear and MLP layers | PVTv2-B0 · ImageNet | latency: 7.82 -> 1.51 ms | Improved |
| Replace dense attention MatMuls with additions and use shifts/MoE for linear and MLP layers | PVTv2-B0 · ImageNet | accuracy: 70.44 -> 70.59 | Improved |
| Replace attention and MLP operations with additions, shifts, and MoE | GNT · LLFF averaged, eight scenes | latency: 1071 -> 746.6 | Improved |
| Replace MLPs with Shift layers rather than using MoE (ablation) | PVTv1-T · ImageNet sensitivity analysis | accuracy: 76.21 -> 73.92 | Hurt |
| Use MoE instead of Shift-only MLPs (ablation) | PVTv1-T · ImageNet sensitivity analysis | accuracy: 73.92 -> 74.81 | Improved |
| Use vanilla Q/K quantization instead of KSH (ablation) | ShiftAddViT · ImageNet classification | accuracy: ↓0.05% ~ ↑0.21% | Mixed |
| Add latency-aware load-balancing loss to MoE routing (ablation) | ShiftAddViT · PVTv2-B0 and PVTv1-T | normalized latency: 100% -> 85.4%; 100% -> 85.5% | Improved |
| Add latency-aware load-balancing loss to MoE routing (ablation) | ShiftAddViT · PVTv2-B0 and PVTv1-T | accuracy: 70.38 -> 70.37; 74.73 -> 74.66 | No clear effect |
| Replace linear/MLP layers with shift layers (ablation) | ShiftAddViT · LLFF novel view synthesis | PSNR: ↑0.13 ~ ↑0.20 | Improved |
| Use linear attention instead of MSA (ablation) | PVTv2-B0 · ImageNet, batch size 1, input resolution 224 | latency: 4.62 -> 6.34 ms | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Ecoformer — EcoFormer: Energy-Saving Attention with Linear Complexity (NeurIPS 2022) · better on 3 of 3 shared results
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
- EfficientViT: Memory Efficient Vision Transformer With Cascaded Group Attention (CVPR 2023 · similarity 0.89)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.88)
- HiRED: Attention-Guided Token Dropping for Efficient Inference of High-Resolution Vision-Language Models (AAAI 2025 · similarity 0.88)
- Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models (NeurIPS 2024 · similarity 0.88)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.87)
- SHViT: Single-Head Vision Transformer with Memory Efficient Macro Design (CVPR 2024 · similarity 0.87)
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.
