Research map / Architectures and efficiency / RNNs and normalization
Towards Adaptive Residual Network Training: A Neural-ODE Perspective
Chengyu Dong, Liyuan Liu, Zichao Li, Jingbo Shang
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 |
|---|---|---|---|
| Adaptive depth growth from shallow start (LipGrow) instead of training a deep ResNet from scratch | ResNet-50 and ResNet-74 on CIFAR-10 · CIFAR-10 | Test accuracy: -0.60 (ResNet-50), -0.30 (ResNet-74) | Mixed |
| Adaptive depth growth from shallow start (LipGrow) | ResNet-50 on CIFAR-100 · CIFAR-100 | Test accuracy: +0.30 | Improved |
| Adaptive depth growth from shallow start (LipGrow) | ResNet-74 on CIFAR-100 · CIFAR-100 | Test accuracy: -0.41 | Hurt |
| Adaptive depth growth from shallow start (LipGrow) | ResNet-66 on Tiny-ImageNet · Tiny-ImageNet | Test accuracy: +0.69 | Improved |
| Adaptive depth growth from shallow start (LipGrow) reducing parameters per epoch | ResNet on CIFAR and Tiny-ImageNet · CIFAR-10/100 and Tiny-ImageNet | Parameters per epoch (training compute proxy): roughly halved (about 50% reduction claimed) | Improved |
| Clone-based initialization with implicit step-size scaling plus Lipschitz-triggered growth vs. hand-tuned schedule (ablation) | LipGrow vs Hand-Tuned growing schedule · CIFAR-10 | Validation accuracy vs PPE | Mixed |
| Choose first growth epoch via Lipschitz-based LipGrow rather than grid search (ablation) | ResNet-50 first growth timing · CIFAR-10 and CIFAR-100 | Min validation error rate | No clear effect |
| Adaptive cosine annealing learning-rate scheduler vs standard cosine and cosine with restarts (ablation) | growing training schedule · CIFAR (appendix) | Accuracy | No clear effect |
Closest papers by meaning
- Deep Linear Network Training Dynamics from Random Initialization: Data, Width, Depth, and Hyperparameter Transfer (ICML 2025 · similarity 0.89)
- Scaling Properties of Deep Residual Networks (ICML 2021 · similarity 0.88)
- Feature Learning Dynamics in Infinite-Depth Neural Networks (NeurIPS 2026 · similarity 0.87)
- Depthwise Hyperparameter Transfer in Residual Networks: Dynamics and Scaling Limit (ICLR 2024 · similarity 0.87)
- Implicit regularization of deep residual networks towards neural ODEs (ICLR 2024 · similarity 0.87)
- On the Weight Dynamics of Deep Normalized Networks (ICML 2024 · similarity 0.85)
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.
