Research map / Architectures and efficiency / RNNs and normalization
High-Performance Large-Scale Image Recognition Without Normalization
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 unit-wise adaptive gradient clipping | NF-ResNet · ImageNet, large-batch training | Top-1 accuracy | Helped |
| Use LARS instead of AGC | NF-ResNet · ImageNet, large-batch training | Top-1 accuracy | Hurt |
| Exclude the final linear classifier layer from AGC (ablation) | NF-ResNet · ImageNet, batch size 4096 | Training stability | Helped |
| Clip weights in all four residual stages (ablation) | NF-ResNet · ImageNet, batch size 4096 | Training stability | Helped |
| Increase RandAugment magnitude too far relative to image resolution | NFNet · ImageNet | Training stability | Hurt |
| Use modified stage widths (ablation) | NFNet-F0 · ImageNet | Top-1 accuracy: 80.4 -> 80.9 | Helped |
| Add a second 3×3 grouped convolution (ablation) | NFNet-F0 · ImageNet | Top-1 accuracy: 80.9 -> 81.3 | Helped |
| Add MixUp (ablation) | NFNet-F0 · ImageNet | Top-1 accuracy: 81.3 -> 82.2 | Helped |
| Add RandAugment (ablation) | NFNet-F0 · ImageNet | Top-1 accuracy: 82.2 -> 83.2 | Helped |
| Add CutMix (ablation) | NFNet-F0 · ImageNet | Top-1 accuracy: 83.2 -> 83.6 | Helped |
| Use the default ResNet stage widths instead of the modified widths (ablation) | NFNet-F0 · ImageNet | Top-1 accuracy: 83.6 -> 83.1 | Hurt |
| Replace the Normalizer-Free setup with batch normalization (ablation) | NFNet-F0 · ImageNet | Top-1 accuracy: 83.6 -> 83.4 | Hurt |
Closest papers by meaning
- Exploring Normalization in Deep Residual Networks with Concatenated Rectified Linear Units (AAAI 2017 · similarity 0.89)
- Demystifying Batch Normalization in ReLU Networks: Equivalent Convex Optimization Models and Implicit Regularization (ICLR 2022 · similarity 0.89)
- Kalman Normalization: Normalizing Internal Representations Across Network Layers (NeurIPS 2018 · similarity 0.88)
- TaskNorm: Rethinking Batch Normalization for Meta-Learning (ICML 2020 · similarity 0.88)
- Characterizing signal propagation to close the performance gap in unnormalized ResNets (ICLR 2021 · similarity 0.88)
- Deconstructing the Regularization of BatchNorm (ICLR 2021 · similarity 0.88)
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.