Research map / Architectures and efficiency / RNNs and normalization
Evolving Normalization-Activation Layers
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 |
|---|---|---|---|
| Use EvoNorm-B0 instead of BN-SiLU/Swish | ImageNet ResNet-50 · ImageNet, ResNet-50 | accuracy: 76.6 ± 0.1 -> 76.6 ± 0.0 | No clear effect |
| Use EvoNorm-B0 instead of BN-SiLU/Swish | ImageNet MobileNetV2 · ImageNet, MobileNetV2 | accuracy: 74.5 ± 0.1 -> 75.0 ± 0.1 | Helped |
| Use EvoNorm-S0 instead of GN-SiLU/Swish | ImageNet ResNet-50 with 4 images/worker · ImageNet, ResNet-50, 4 images/worker | accuracy: 76.5 ± 0.0 -> 76.5 ± 0.1 | No clear effect |
| Replace BN-ReLU with EvoNorm-B0 | Mask R-CNN with ResNet-50-FPN · COCO val2017 | bbox AP: 42.1 -> 44.0 (+1.9) | Helped |
| Replace GN-ReLU with EvoNorm-S0 | Mask R-CNN with ResNet-50-FPN · COCO val2017 | bbox AP: 42.7 -> 43.6 (+0.9) | Helped |
| Replace BN-ReLU with EvoNorm-B0 | Mask R-CNN with SpineNet-96 · COCO val2017 | bbox AP: 47.1 -> 48.0 (+0.9) | Helped |
| Replace GN-ReLU with EvoNorm-S0 | Mask R-CNN with SpineNet-96 · COCO val2017 | bbox AP: 45.7 -> 47.5 (+1.8) | Helped |
| Replace BN-ReLU with EvoNorm-B0 | BigGAN-deep · ImageNet generation 128×128 | FID: 7.85/7.29 -> 6.91/5.87 | Helped |
| Replace BN-ReLU with EvoNorm-B0 | BigGAN-deep · ImageNet generation 128×128 | Inception Score: 118.77/124.01 -> 101.13/113.63 | Hurt |
| Add the v1 x term to the local denominator term (ablation) | EvoNorm-B0 · ImageNet, ResNet-50 | accuracy: 14.4; 50.5; 22.0 -> 76.5; 76.6; 76.6 | Helped |
| Add the local denominator term (ablation) | EvoNorm-B0 · ImageNet, ResNet-50 | accuracy: 4.2; 4.1; 4.1 -> 76.5; 76.6; 76.6 | Helped |
| Add the global batch-statistics denominator term (ablation) | EvoNorm-B0 · ImageNet, ResNet-50 | accuracy: 1e-3; 1e-3; 1e-3 -> 76.5; 76.6; 76.6 | Helped |
Closest papers by meaning
- Beyond BatchNorm: Towards a Unified Understanding of Normalization in Deep Learning (NeurIPS 2021 · similarity 0.88)
- LayerAct: Advanced Activation Mechanism for Robust Inference of CNNs (AAAI 2025 · similarity 0.86)
- Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion (ICLR 2024 · similarity 0.85)
- On the Weight Dynamics of Deep Normalized Networks (ICML 2024 · similarity 0.85)
- On the Importance of Gaussianizing Representations (ICML 2025 · similarity 0.84)
- On the Nonlinearity of Layer Normalization (ICML 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.