Research map / Architectures and efficiency / RNNs and normalization
Commutative Scaling of Width and Depth in Deep Neural Networks
Soufiane Hayou
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 |
|---|---|---|---|
| increase width n from 8 to 256 to 16384 | ResNet with uniform scaling alpha=L^-1/2 · Neural covariance at depth L=2 | L2 error to theory: 2.34e-01 -> 1.32e-02 -> 1.04e-03 | Improved |
| increase depth L from 2 to 256 at width 2^14 | ResNet with uniform scaling alpha=L^-1/2 · Neural covariance, width 16384 | L2 error to theory: 1.04e-03 -> 1.62e-04 | Improved |
| increase depth L from 2 to 256 at width 8 | ResNet with uniform scaling alpha=L^-1/2 · Neural covariance, width 8 | L2 error to theory: 2.34e-01 -> 6.08e-01 | Hurt |
| use shaped ReLU MLP instead of scaled ResNet | Proportional width-depth limit L=n · Neural covariance distribution | concentration around sequential limit | Hurt |
| use Shaped ResNet with main branch scaled by beta=1/2 instead of alpha=L^-1/2 | Residual network · Neural covariance distribution, L=n | concentration around sequential limit | Hurt |
| improved depth rate analysis versus prior rate L^-1/2 | Infinite-width covariance for alpha=L^-1/2 · Delta_L vs depth L in 2^3..2^14 | convergence rate in L: L^-1/2 -> L^-1 | Improved |
Closest papers by meaning
- On the Surprising Effectiveness of Large Learning Rates under Standard Width Scaling (NeurIPS 2025 · similarity 0.87)
- Infinite-Width Limit of a Single Attention Layer: Analysis via Tensor Programs (NeurIPS 2025 · similarity 0.86)
- Tensor Programs VI: Feature Learning in Infinite Depth Neural Networks (ICLR 2024 · similarity 0.85)
- Feature Learning Dynamics in Infinite-Depth Neural Networks (NeurIPS 2026 · similarity 0.85)
- Geometric Inductive Biases of Deep Networks: The Role of Data and Architecture (ICLR 2025 · similarity 0.85)
- Super Consistency of Neural Network Landscapes and Learning Rate Transfer (NeurIPS 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.
