Research map / Architectures and efficiency / Mixture of experts
OneSparse: A Unified Framework for Sparse Activation Layers in Vision Models
Xingkui Zhu, Dingkang Liang, Cheng Chen, Daoxin Zhang, lv Hanxiang, Zhe Xu, Yao Hu, Xiang Bai
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 FFN with Nexus Layer (same FLOPs) | ViT-S · ImageNet-1K val | Top-1 accuracy (%): 78.8 -> 81.2 | Improved |
| use Nexus instead of SoftMoE | ViT-S · ImageNet-1K val | Top-1 accuracy (%): 80.8 -> 81.2, FLOPs 5.4G -> 4.3G | Improved |
| use Nexus instead of Memory+ | ViT-S · ImageNet-1K val | Top-1 accuracy (%): 79.3 -> 81.2 | Improved |
| use Nexus instead of SoftMoE | ConvNeXt-S · ImageNet-1K val | Top-1 accuracy (%): 83.9 -> 83.9 | No clear effect |
| use Nexus instead of dense | ConvNeXt-B · ImageNet-1K val | Top-1 accuracy (%): 83.8 -> 84.5 | Improved |
| add Memory+ layers | ConvNeXt-B · ImageNet-1K val | Top-1 accuracy (%): 83.8 -> 83.8 | No clear effect |
| use Nexus instead of dense | ViT-S Mask R-CNN · COCO val2017 | APbbox: 40.2 -> 42.7 | Improved |
| use Nexus instead of dense | ViT-S · ADE20K | mIoU (%): 44.6 -> 45.8 | Improved |
| use unified routing instead of Product Key routing in memory layers (ablation) | ViT-S · ImageNet-1K val | Top-1 accuracy (%): 79.4 -> 80.1 | Improved |
| use unified routing instead of Expert Choice routing over memory values (ablation) | ViT-S · ImageNet-1K val | FLOPs (G): 59.9 -> 4.2 (EC runs out of memory) | Improved |
| add expert units to a memory-only layer under fixed parameter budget (ablation) | Nexus · ImageNet-1K val | Top-1 accuracy (%): 80.1 -> 81.2 | Improved |
| use a hybrid memory/expert allocation instead of expert-only (ablation) | Nexus · ImageNet-1K val | Top-1 accuracy (%): 80.8 -> 81.2, expert-only costs over 1.2x compute | Improved |
Closest papers by meaning
- SparseFormer: Sparse Visual Recognition via Limited Latent Tokens (ICLR 2024 · similarity 0.88)
- From Sparse to Soft Mixtures of Experts (ICLR 2024 · similarity 0.87)
- LookThere! Sparse Vision by Reinforced Selection (NeurIPS 2026 · similarity 0.87)
- CoreMatching: A Co-adaptive Sparse Inference Framework with Token and Neuron Pruning for Comprehensive Acceleration of Vision-Language Models (ICML 2025 · similarity 0.87)
- From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit (NeurIPS 2025 · similarity 0.86)
- LookWhere? Efficient Visual Recognition by Learning Where to Look and What to See from Self-Supervision (NeurIPS 2025 · similarity 0.86)
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.
