Research map / Architectures and efficiency / Attention and transformers
GOHSP: A Unified Framework of Graph and Optimization-Based Heterogeneous Structured Pruning for Vision Transformer
Miao Yin, Burak Uzkent, Yilin Shen, Hongxia Jin, Bo Yuan
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 |
|---|---|---|---|
| Use graph-ranked head importance as a soft mask and jointly optimize heterogeneous structured sparsity rather than pruning heads and columns directly | ViT-Small · CIFAR-10, 40% sparsity | Top-1 accuracy: 96.93 -> 97.89 | Improved |
| Use GOHSP structured pruning instead of SSP | DeiT-Tiny · ImageNet, 30% sparsity | Top-1 accuracy: +1.65% | Improved |
| Use GOHSP structured pruning instead of S²ViTE | DeiT-Small · ImageNet, 40% sparsity | Top-1 accuracy: +0.76% | Improved |
| Compress ViT-Small with GOHSP at 40% sparsity | ViT-Small · CIFAR-10 | Top-1 accuracy: 97.85 -> 97.89 | Improved |
| Increase GOHSP sparsity from 40% to 80% | ViT-Small · CIFAR-10 | Top-1 accuracy: 97.89 -> 97.40 | Hurt |
| Increase GOHSP sparsity from 40% to 50% | DeiT-Small · ImageNet | Top-1 accuracy: 79.98 -> 79.86 | Hurt |
| Use soft pruning instead of directly hard-pruning the lowest-ranked heads and lowest-norm MLP columns (ablation) | ViT-Small · CIFAR-10 | Top-1 accuracy | Improved |
| Estimate head ranking from a sampled batch rather than the full training dataset (ablation) | GOHSP head-ranking algorithm · ViT-Small on CIFAR-10 | Head-ranking stability | No clear effect |
| Increase the penalty parameter ρ (ablation) | GOHSP · ViT-Small on CIFAR-10 | Sparsity imposed during training | Improved |
| Set the penalty parameter ρ too high (ablation) | GOHSP · ViT-Small on CIFAR-10 | Top-1 accuracy | Hurt |
Later papers that compared with this method
2 later accepted papers report GOHSP as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Adaptive Pruning of Pretrained Transformer via Differential Inclusions (ICLR 2025 · better on 2 of 2 shared results)
- Isomorphic Pruning for Vision Models (ECCV 2024 · better on 1 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
- MDP: Multidimensional Vision Model Pruning with Latency Constraint (CVPR 2025 · similarity 0.90)
- Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers (CVPR 2024 · similarity 0.89)
- UPDP: A Unified Progressive Depth Pruner for CNN and Vision Transformer (AAAI 2024 · similarity 0.88)
- Collaborative Multi-Mode Pruning for Vision-Language Models (CVPR 2026 · similarity 0.88)
- Data-independent Module-aware Pruning for Hierarchical Vision Transformers (ICLR 2024 · similarity 0.88)
- CORP: Closed-Form One-shot Representation-Preserving Structured Pruning for Transformers (NeurIPS 2026 · similarity 0.88)
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.
