Research map / Architectures and efficiency / Pruning and efficient tuning
Kernelized Sparse Fine-Tuning With Bi-Level Parameter Competition for Vision Models
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 linear kernel with non-linear P-Linear kernel (ablation) | Kernelized LoRA (no sparsification) · VTAB-1k ViT-B/16 | mean accuracy: 75.2% v.s. 74.2% | Improved |
| add sparsification to a linear-kernel LoRA (ablation) | Linear kernel LoRA · VTAB-1k ViT-B/16 | mean accuracy: 74.2% v.s. 65.7% | Hurt |
| add sparsification to a P-Linear kernel LoRA (ablation) | P-Linear kernelized LoRA · VTAB-1k ViT-B/16 | mean accuracy: 75.2% v.s. 76.9% | Improved |
| use Mix-K instead of P-Linear kernel (with sparsification) (ablation) | SNELLA · VTAB-1k ViT-B/16 | mean accuracy: 77.2% v.s. 76.9% | Improved |
| use Mix-K instead of P-Linear kernel (no sparsification) (ablation) | Kernelized LoRA · VTAB-1k ViT-B/16 | mean accuracy: 75.7% v.s. 75.2% | Improved |
| normalize RBF kernel values (ablation) | RBF kernelized LoRA · VTAB-1k ViT-B/16 | mean accuracy: 75.2% v.s. 58.7% | Improved |
| use Mix-K instead of normalized RBF kernel (ablation) | Kernelized LoRA · VTAB-1k ViT-B/16 | mean accuracy: 75.7% v.s. 75.2% | Improved |
| add within-layer competition sparsity (ablation) | Kernelized LoRA · VTAB-1k ViT-B/16 | mean accuracy: 75.7% v.s. 77.0% | Improved |
| add across-layer competition (ablation) | KLoRA with within-layer competition · VTAB-1k ViT-B/16 | mean accuracy: 77.0% v.s. 77.2% | Improved |
| use sensitivity instead of magnitude as layer importance (alloc per epoch) (ablation) | SNELLA · VTAB-1k ViT-B/16 (0.01% params) | mean accuracy: 75.1 -> 75.3 | Improved |
| reallocate sparsity per epoch instead of per step (sensitivity) (ablation) | SNELLA · VTAB-1k ViT-B/16 (0.01% params) | mean accuracy: 75.0 -> 75.3 | Improved |
| add budget decay scheduling (linear) vs constant budget (ablation) | SNELLA-32 · VTAB-1k 4 datasets | mean accuracy: 61.07% v.s. 63.50% | Improved |
Closest papers by meaning
- Expanding Sparse Tuning for Low Memory Usage (NeurIPS 2024 · similarity 0.96)
- Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation (CVPR 2025 · similarity 0.89)
- Unleashing the Power of Meta-tuning for Few-shot Generalization Through Sparse Interpolated Experts (ICML 2024 · similarity 0.88)
- PELA: Learning Parameter-Efficient Models with Low-Rank Approximation (CVPR 2024 · similarity 0.87)
- SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors (NeurIPS 2024 · similarity 0.87)
- Lessons and Insights from a Unifying Study of Parameter-Efficient Fine-Tuning (PEFT) in Visual Recognition (CVPR 2025 · 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.
