Research map / Architectures and efficiency / Pruning and efficient tuning
CaRe: Critical Parameter Rectification for Efficient Visual Modeling
Mingjia Li, Hongkun Xiong, Yuheng Shi, Hengxing Liu, Xiaojie Guo
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 Softplus instead of identity as gate activation (ablation) | CaReNet-M1 · ImageNet-1K | top-1 accuracy (%): 72.9 -> 73.6 | Improved |
| use GELU instead of identity as gate activation (ablation) | CaReNet-M1 · ImageNet-1K | top-1 accuracy (%): 72.9 -> 73.5 | Improved |
| add a dense base weight to the gated term (ablation) | CaRe (W = P1 ⊙ φ(P2)) · ImageNet-1K | top-1 accuracy (%): 69.5 -> 73.6 | Improved |
| gate the single weight with no base weight (vs standard conv) (ablation) | standard convolution · ImageNet-1K | top-1 accuracy (%): 72.7 -> 69.5 | Hurt |
| keep the gated auxiliary term at inference (vs discarding after training) (ablation) | CaReNet-M1 · ImageNet-1K | top-1 accuracy (%): 60.1 -> 73.6 | Improved |
| use learned gated correction instead of RepVGG-style fixed correction (ablation) | CaReNet-M1 · ImageNet-1K | top-1 accuracy (%): 73.2 -> 73.6 | Improved |
| use two gated terms instead of one gated plus one dense term (ablation) | CaReNet-M1 · ImageNet-1K | top-1 accuracy (%): 73.6 -> 70.1 | Hurt |
| use Progressive fusion instead of final-stage-only fusion (ablation) | CaReNet-M1 · ImageNet-1K | top-1 accuracy (%): 73.1 -> 73.6 | Improved |
| use Progressive fusion instead of uniform-average fusion (ablation) | CaReNet-M1 · ImageNet-1K | top-1 accuracy (%): 72.9 -> 73.6 | Improved |
| add CaRe to ResNet-18 | ResNet-18 · ImageNet-1K (DeiT recipe) | top-1 accuracy (%): 69.1 -> 69.3 | Improved |
| add CaRe to ViT-Tiny | ViT-Tiny · ImageNet-1K (DeiT recipe) | top-1 accuracy (%): 72.2 -> 72.4 | Improved |
| add CaRe to PoolFormerV2-S12 | PoolFormerV2-S12 · ImageNet-1K (DeiT recipe) | top-1 accuracy (%): 77.7 -> 78.1 | Improved |
Closest papers by meaning
- RePaViT: Scalable Vision Transformer Acceleration via Structural Reparameterization on Feedforward Network Layers (ICML 2025 · similarity 0.85)
- Densely Connected Parameter-Efficient Tuning for Referring Image Segmentation (AAAI 2025 · similarity 0.85)
- Make RepVGG Greater Again: A Quantization-Aware Approach (AAAI 2024 · similarity 0.85)
- PELA: Learning Parameter-Efficient Models with Low-Rank Approximation (CVPR 2024 · similarity 0.84)
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 2025 · similarity 0.83)
- Scaling Exponents Across Parameterizations and Optimizers (ICML 2024 · similarity 0.83)
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.
