Research map / Architectures and efficiency / RNNs and normalization
Mode Normalization
Lucas Deecke, Iain Murray, Hakan Bilen
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 |
|---|---|---|---|
| replace batch normalization with mode normalization | LeNet · MNIST, CIFAR-10, SVHN, Fashion-MNIST multi-task; N=128 | test error (%): 26.91 ± 1.08 -> 23.16 ± 1.23 | Improved |
| use mode group normalization instead of group normalization | LeNet · MNIST, CIFAR-10, SVHN, Fashion-MNIST multi-task; N=4, 8, and 16 | test error (%): 32.15 ± 1.10 -> 31.30 ± 1.65; 28.60 ± 1.45 -> 26.83 ± 1.34; 27.63 ± 0.45 -> 26.00 ± 1.68 | Improved |
| replace batch normalization with mode normalization | NIN · CIFAR-10 | test error (%): 8.82 -> 8.42 | Improved |
| replace batch normalization with mode normalization | NIN · CIFAR-100 | test error (%): 32.30 -> 31.66 | Improved |
| replace batch normalization with mode normalization | VGG13 · CIFAR-10 | test error (%): 8.28 -> 7.79 | Improved |
| replace batch normalization with mode normalization | VGG13 · CIFAR-100 | test error (%): 31.15 -> 30.06 | Improved |
| replace batch normalization with mode normalization | ResNet20 · CIFAR-10 | test error (%): 8.44 -> 7.99 | Improved |
| replace batch normalization with mode normalization | ResNet20 · CIFAR-100 | test error (%): 31.56 -> 30.53 | Improved |
| replace batch normalization with mode normalization | ResNet18 · ILSVRC12 | top-1 error (%): 30.25 -> 30.07 | Improved |
| replace batch normalization with mode normalization | ResNet18 · ILSVRC12 | top-5 error (%): 10.90 -> 10.65 | Improved |
| increase the number of modes from K=2 to K=4 (ablation) | LeNet with mode normalization · MNIST, CIFAR-10, SVHN, Fashion-MNIST multi-task; N=128 | test error (%): 23.16 ± 1.23 -> 24.25 ± 0.71 | Hurt |
| increase the number of modes from K=2 to K=6 (ablation) | LeNet with mode normalization · MNIST, CIFAR-10, SVHN, Fashion-MNIST multi-task; N=128 | test error (%): 23.16 ± 1.23 -> 25.12 ± 1.48 | Hurt |
Closest papers by meaning
- Training Faster by Separating Modes of Variation in Batch-Normalized Models (TPAMI 2019 · similarity 0.91)
- Normalization Propagation: A Parametric Technique for Removing Internal Covariate Shift in Deep Networks (ICML 2016 · similarity 0.88)
- Towards Understanding Regularization in Batch Normalization (ICLR 2019 · similarity 0.88)
- Four Things Everyone Should Know to Improve Batch Normalization (ICLR 2020 · similarity 0.88)
- Understanding Batch Normalization (NeurIPS 2018 · similarity 0.87)
- Batch Normalization Is Blind to the First and Second Derivatives of the Loss (AAAI 2024 · similarity 0.86)
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.
