Research map / Architectures and efficiency / RNNs and normalization
Maximal Initial Learning Rates in Deep ReLU Networks
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 a sufficiently large width-to-depth ratio and train the input layer at a relatively small learning rate | Constant-width fully connected ReLU networks · MNIST and CIFAR-10 | maximal initial learning rate η* scaling with depth × width | Helped |
| Train the input layer at the same learning rate as the other layers (ablation) | Constant-width fully connected ReLU networks · CIFAR-10 and MNIST | linearity of η* versus depth × width | Hurt |
| Freeze the input layer at initialization (ablation) | Constant-width fully connected ReLU networks · Not specified | η* power-law trend | No clear effect |
| Increase the threshold accuracy (ablation) | Constant-width fully connected ReLU networks with sufficiently large width/depth · MNIST | maximal initial learning rate η* | No clear effect |
| Increase the threshold accuracy (ablation) | Constant-width fully connected ReLU networks with smaller width/depth · MNIST | maximal initial learning rate η* | Hurt |
| Use NTK parametrization instead of Kaiming initialization (ablation) | Fully connected ReLU networks · CIFAR-10 | relationship between 2/λ₁ and η* | Mixed |
| Use LeCun initialization instead of Kaiming initialization (ablation) | Fully connected ReLU networks · Initialization sharpness | sharpness λ₁ scaling with depth | Mixed |
| Train at η* rather than use a well-tuned constant learning rate (ablation) | ResNet-20 · Not specified | performance | Hurt |
| Use a learning-rate scheduler with η* (ablation) | ResNet-20 · Not specified | performance | Helped |
| Train with η* rather than a smaller fixed learning rate | Fully connected networks · CIFAR-10 validation | accuracy | Mixed |
| Apply the condition η ≤ 2/λ₁ as a necessary constraint at initialization | Kaiming-initialized fully connected ReLU networks · MNIST and CIFAR-10 | ability to reach threshold accuracy | Hurt |
Closest papers by meaning
- Optimal Rates for Generalization of Gradient Descent for Deep ReLU Classification (NeurIPS 2025 · similarity 0.88)
- Compelling ReLU Networks to Exhibit Exponentially Many Linear Regions at Initialization and During Training (ICML 2025 · similarity 0.88)
- Where Do Large Learning Rates Lead Us? (NeurIPS 2024 · similarity 0.87)
- Deep Linear Network Training Dynamics from Random Initialization: Data, Width, Depth, and Hyperparameter Transfer (ICML 2025 · similarity 0.87)
- On the Surprising Effectiveness of Large Learning Rates under Standard Width Scaling (NeurIPS 2025 · similarity 0.87)
- Simplicity Bias and Optimization Threshold in Two-Layer ReLU Networks (ICML 2025 · similarity 0.87)
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.