Research map / Architectures and efficiency / RNNs and normalization
PWLU: Learning Specialized Activation Functions With the Piecewise Linear Unit
Zezhou Zhu, Yucong Zhou, Yuan Dong, Zhao Zhong
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 |
|---|---|---|---|
| add statistic-based boundary realignment instead of fixed init (x=10) (ablation) | channel-wise PWLU, ResNet-18 · ImageNet-1k val | top-1 accuracy (%): 72.19 -> 72.57 | Improved |
| add statistic-based boundary realignment instead of fixed init (x=10) (ablation) | channel-wise PWLU, MobileNet-V3 · ImageNet-1k val | top-1 accuracy (%): 68.20 -> 69.65 | Improved |
| add statistic-based realignment instead of fixed init (x=10) (ablation) | layer-wise PWLU, ResNet-18 · ImageNet-1k val | top-1 accuracy (%): 71.84 -> 72.22 | Improved |
| add statistic-based realignment instead of fixed init (x=3) (ablation) | layer-wise PWLU, MobileNet-V3 · ImageNet-1k val | top-1 accuracy (%): 68.54 -> 69.30 | Improved |
| use channel-wise instead of layer-wise PWLU (ablation) | PWLU stat-realign, ResNet-18 · ImageNet-1k val | top-1 accuracy (%): 72.22 -> 72.57 | Improved |
| use 8 intervals instead of 4 (ablation) | channel-wise PWLU, MobileNet-V3 · ImageNet-1k val | top-1 accuracy (%): 68.70 -> 69.55 | Improved |
| use 16 intervals instead of 4 (ablation) | channel-wise PWLU, ResNet-18 · ImageNet-1k val | top-1 accuracy (%): 72.24 -> 72.57 | Improved |
| use 20 intervals instead of 16 (ablation) | channel-wise PWLU, MobileNet-V3 · ImageNet-1k val | top-1 accuracy (%): 69.65 -> 69.37 | Hurt |
| use 20 intervals instead of 16 (ablation) | channel-wise PWLU, EfficientNet-B0 · ImageNet-1k val | top-1 accuracy (%): 78.22 -> 77.86 | Hurt |
| replace ReLU with PWLU | MobileNet-V3 · ImageNet-1k val | top-1 accuracy (%): 67.74 -> 69.65 | Improved |
| replace ReLU with PWLU | ResNet-18 · ImageNet-1k val | top-1 accuracy (%): 71.34 -> 72.57 | Improved |
| replace Swish with PWLU | Mask R-CNN ResNet-50-FPN · COCO val2017 | box AP: 40.05 -> 40.48 | Improved |
Closest papers by meaning
- Learning Specialized Activation Functions With the Piecewise Linear Unit (ICCV 2021 · similarity 0.98)
- Smooth Maximum Unit: Smooth Activation Function for Deep Networks Using Smoothing Maximum Technique (CVPR 2022 · similarity 0.87)
- SAU: Smooth Activation Function Using Convolution with Approximate Identities (ECCV 2022 · similarity 0.85)
- Gompertz Linear Units: Leveraging Asymmetry for Enhanced Learning Dynamics (NeurIPS 2025 · similarity 0.84)
- ReLUs Are Sufficient for Learning Implicit Neural Representations (ICML 2024 · similarity 0.83)
- Better NTK Conditioning: A Free Lunch from (ReLU) Nonlinear Activation in Wide Neural Networks (NeurIPS 2025 · 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.
