Research map / Optimization / Implicit bias and generalization
Training Scale-Invariant Neural Networks on the Sphere Can Happen in Three Regimes
Maxim Kodryan, Ekaterina Lobacheva, Maksim Nakhodnov, Dmitry P. Vetrov
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 |
|---|---|---|---|
| Train with increasing fixed effective learning rates on the sphere | Scale-invariant neural networks trained with projected SGD · CIFAR-10 | training loss | Mixed |
| Increase the ELR while remaining in the convergence regime | Scale-invariant neural networks trained on the sphere · CIFAR-10 | test error | Improved |
| Increase the ELR while remaining in the convergence regime | Scale-invariant neural networks trained on the sphere · CIFAR-10 | sharpness | Improved |
| Fine-tune with a decreased ELR | A converged scale-invariant neural network · CIFAR-10 | training trajectory | No clear effect |
| Fine-tune with an increased ELR | A converged scale-invariant neural network · CIFAR-10 | sharpness | Improved |
| Increase ELR in the chaotic-equilibrium regime | Scale-invariant neural networks trained on the sphere · CIFAR-10 | training loss | Hurt |
| Use low ELRs from the second regime before training with first-regime ELRs | Scale-invariant neural networks trained on the sphere · CIFAR-10 | sharpness | Improved |
| Use a medium starting learning rate and cosine decay | ResNet-18 with conventional training · CIFAR-10 | test accuracy | Improved |
| Use a high starting learning rate and cosine decay | ResNet-18 with conventional training · CIFAR-10 | training loss | Hurt |
| Add data augmentation | Scale-invariant ConvNet · CIFAR-10 | training loss | Hurt |
| Increase network width | Scale-invariant ConvNet · CIFAR-10 | training loss | Mixed |
| Add label noise | Scale-invariant ConvNet · CIFAR-10 | sharpness | Hurt |
Closest papers by meaning
- On the Weight Dynamics of Deep Normalized Networks (ICML 2024 · similarity 0.87)
- Loss Landscape Characterization of Neural Networks without Over-Parametrization (NeurIPS 2024 · similarity 0.86)
- Deep Linear Network Training Dynamics from Random Initialization: Data, Width, Depth, and Hyperparameter Transfer (ICML 2025 · similarity 0.86)
- On the Surprising Effectiveness of Large Learning Rates under Standard Width Scaling (NeurIPS 2025 · similarity 0.86)
- Scalable Optimization in the Modular Norm (NeurIPS 2024 · similarity 0.86)
- Learning in Compact Spaces with Approximately Normalized Transformer (NeurIPS 2025 · 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.
