Research map / Architectures and efficiency / RNNs and normalization
Module-wise Training of Neural Networks via the Minimizing Movement Scheme
Skander Karkar, Ibrahim Ayed, Emmanuel de Bezenac, patrick gallinari
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 transport regularization (TRGL vs VanGL) (ablation) | parallel module-wise training, ResNet-110, 16 modules · STL10 | test accuracy (%): 73.57 -> 77.22 | Improved |
| add transport regularization (TRGL vs VanGL) (ablation) | parallel module-wise training, ResNet-110, 4 modules · STL10 | test accuracy (%): 77.11 -> 79.72 | Improved |
| add transport regularization (TRGL vs VanGL) (ablation) | parallel, ResNet-101, 4 modules · TinyImageNet | test accuracy (%): 63.64 -> 63.71 | No clear effect |
| add transport regularization (TRGL vs VanGL) (ablation) | parallel, Swin-Tiny, 4 modules · CIFAR100 | test accuracy (%): 69.34 -> 74.11 | Improved |
| add transport regularization (TRGL vs VanGL) (ablation) | parallel, 2-7 ResNet · CIFAR10 | test accuracy (%): 94.01 -> 94.05 | No clear effect |
| add transport regularization (ablation) | sequential block-wise 10-1 ResNet · CIFAR10, train size 1000 | last block test accuracy (%): 45.67 -> 51.34 | Improved |
| add transport regularization (ablation) | sequential block-wise 10-1 ResNet · CIFAR100, train size 50000 | highest test accuracy (%): 68.74 -> 68.79 | No clear effect |
| add transport regularization (ablation) | parallel 10-1 ResNet · CIFAR100, train size 5000 | highest test accuracy (%): 39.45 -> 40.36 | Improved |
| use multi-lap sequential training instead of sequential (ablation) | TRGL, 10-1 ResNet · CIFAR100, train size 50000 | highest test accuracy (%): 68.79 -> 69.95 | Improved |
| use multi-lap sequential training instead of sequential (ablation) | VanGL, 10-1 ResNet · CIFAR100, train size 5000 | highest test accuracy (%): 36.37 -> 33.68 | Hurt |
| add multi-lap sequential training to sequential (ablation) | ResNet-101 split into 2 modules · CIFAR100 | test accuracy (%): 73.31 -> 74.34 | Improved |
| add regularization to multi-lap sequential (ablation) | ResNet-101 split into 2 modules · CIFAR100 | test accuracy (%): 74.34 -> 74.78 | Improved |
Closest papers by meaning
- MinMax Methods for Optimal Transport and Beyond: Regularization, Approximation and Numerics (NeurIPS 2020 · similarity 0.85)
- Train Faster, Perform Better: Modular Adaptive Training in Over-Parameterized Models (NeurIPS 2023 · similarity 0.85)
- Learning Sparse Neural Networks through L_0 Regularization (ICLR 2018 · similarity 0.84)
- Implicit Self-Regularization in Deep Neural Networks: Evidence from Random Matrix Theory and Implications for Learning (JMLR 2021 · similarity 0.84)
- Expectile Regularization for Fast and Accurate Training of Neural Optimal Transport (NeurIPS 2024 · similarity 0.84)
- Deep Linear Network Training Dynamics from Random Initialization: Data, Width, Depth, and Hyperparameter Transfer (ICML 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.
