Research map / Architectures and efficiency / Pruning and efficient tuning
LoPrune: Efficient Data Pruning for LoRA-Based Fine-Tuning of Vision Transformer
Qiang He, Yaozong Yang, Kaibin Wang, Ziteng Wei, Feifei Chen, Caslon Chua, Yun Yang
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 |
|---|---|---|---|
| project influence scores onto the LoRA trainable subspace (TSAS) (ablation) | LoPrune (K-FAC) · CIFAR-100 | accuracy: 86.88 -> 89.59 | Improved |
| project influence scores onto the LoRA trainable subspace (TSAS) (ablation) | LoPrune (K-FAC) · Tiny-ImageNet | accuracy: 82.45 -> 84.78 | Improved |
| project influence scores onto the LoRA trainable subspace (TSAS) (ablation) | LoPrune · CIFAR-100 (ViT-Small, pruning ratio 0.8) | accuracy: 80.73 -> 88.65 | Improved |
| use K-FAC instead of pseudoinverse Hessian (ablation) | LoPrune · CIFAR-100 | total time (s): 5542.3 -> 371.5 | Improved |
| use K-FAC instead of pseudoinverse Hessian (ablation) | LoPrune · CIFAR-100 | accuracy: 89.64 -> 89.59 | No clear effect |
| use K-FAC instead of the Agarwal approximation (ablation) | LoPrune · CIFAR-100 | accuracy: 89.13 -> 89.59 | Improved |
| start pruning later in fine-tuning (fully converged model for scoring) (ablation) | LoPrune · ViT-Small, CIFAR-100 and Tiny-ImageNet | accuracy: max +0.22% accuracy, training time up to +68% | No clear effect |
| LoPrune vs random pruning at 0.5 on COCO/classification (replace random with TSA scoring) | DeiT-Tiny · CIFAR-100 (0.5 pruning) | top-1 accuracy: 77.20 -> 79.88 | Improved |
| use warm-up-based scoring baselines (EL2N, GraNd) with few fine-tuning epochs | EL2N / GraNd · CIFAR-100 and Tiny-ImageNet, high pruning ratio | accuracy: 5.08% - 10.34% lower than LoPrune above 0.8 pruning | Hurt |
| prune a small fraction (10%) of data with various pruning methods | ViT-Small fine-tuning · CIFAR-100 and Tiny-ImageNet | accuracy | Improved |
| fine-tune ResNets on a subset selected by LoPrune with ViT-Small instead of random subset | ResNet-50 · CIFAR-100 (0.8 ratio retained) | top-1 accuracy: 74.38 -> 77.83 | Improved |
| LoPrune at 0.5 pruning on Tiny-ImageNet | ViT-Small · Tiny-ImageNet | total training time (s): 992.1 -> 622.1 | Improved |
Closest papers by meaning
- RandLoRA: Full rank parameter-efficient fine-tuning of large models (ICLR 2025 · similarity 0.86)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.85)
- Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers (CVPR 2024 · similarity 0.84)
- Sparse High Rank Adapters (NeurIPS 2024 · similarity 0.84)
- $\textit{Trans-LoRA}$: towards data-free Transferable Parameter Efficient Finetuning (NeurIPS 2024 · similarity 0.84)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · similarity 0.84)
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.
