Research map / Architectures and efficiency / RNNs and normalization
Tanh Works Better with Asymmetry
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| swap order: put BN after Tanh (ablation) | Tanh VGG16_11 · CIFAR-100 | accuracy (%): 69.5 -> 74.11 | Helped |
| swap order: put BN after Tanh | Tanh VGG16 · CIFAR-100 | accuracy (%): 64.95±0.31 -> 70.93±0.12 | Helped |
| swap order: put BN after Tanh | Tanh PreAct-ResNet18 · CIFAR-10 | accuracy (%): 92.46±0.17 -> 94.41±0.15 | Helped |
| swap order: put BN after Tanh | Tanh MobileNet · ImageNet | accuracy (%): 64.26 -> 72.07 | Helped |
| swap order: put BN after ReLU | ReLU VGG16 · CIFAR-100 | accuracy (%): 73.68±0.08 -> 71.79±0.09 | Hurt |
| swap order: put BN after Shifted Tanh | Shifted Tanh VGG16 · CIFAR-100 | accuracy (%): 73.87±0.05 -> 73.21±0.31 | No clear effect |
| replace Tanh with Shifted Tanh (tau=-1) in Convention order | VGG16 · CIFAR-100 | accuracy (%): 64.95±0.31 -> 73.87±0.05 | Helped |
| swap order on Softsign | Softsign VGG16_11 · CIFAR-100 | accuracy (%): 70.01 -> 73.65 | Helped |
| swap order on LeCun Tanh | LeCun Tanh VGG16_11 · CIFAR-100 | accuracy (%): 67.82 -> 74.46 | Helped |
| encourage asymmetric saturation by removing weight decay on BN affine params (NWDBN) (ablation) | Tanh Convention VGG16_11 · CIFAR-100 | accuracy (%): 69.5 -> 72.22 | Helped |
| swap order on corrupted inputs | Tanh VGG16_11 · CIFAR-100-C | average accuracy (%): 41.4 -> 45.3 | Helped |
| shifted Tanh offset tau from -1.0 to -1.5 (ablation) | Shifted Tanh · CIFAR-10/100 | accuracy (%): 84.15 -> 80.56 | Hurt |
Closest papers by meaning
- Understanding Batch Normalization (NeurIPS 2018 · similarity 0.84)
- Better NTK Conditioning: A Free Lunch from (ReLU) Nonlinear Activation in Wide Neural Networks (NeurIPS 2025 · similarity 0.83)
- Gompertz Linear Units: Leveraging Asymmetry for Enhanced Learning Dynamics (NeurIPS 2025 · similarity 0.82)
- Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion (ICLR 2024 · similarity 0.82)
- Unraveling Batch Normalization for Realistic Test-Time Adaptation (AAAI 2024 · similarity 0.82)
- Batch Normalization Is Blind to the First and Second Derivatives of the Loss (AAAI 2024 · similarity 0.82)
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.