Research map / Architectures and efficiency / CNN design
Finite Versus Infinite Neural Networks: an Empirical Study
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 NNGP instead of NTK inference | Kernel models across FC, CNN-VEC, and CNN-GAP · CIFAR-10, CIFAR-100, and Fashion-MNIST | accuracy | Helped |
| Ensemble kernel predictors over augmented training sets | NNGP and NTK kernels · CIFAR-10 | accuracy: 83.5 -> 84.8 | Helped |
| Apply tuned regularized ZCA preprocessing | CNN-GAP kernels · CIFAR-10 | accuracy: 78.5 -> 83.5 | Helped |
| Use regularized ZCA rather than pure ZCA whitening (ablation) | Finite networks and kernel methods · Image classification | accuracy | Mixed |
| Center the network by subtracting its initial predictions (ablation) | FCN and CNN-VEC · CIFAR-10 | accuracy: 47.82 -> 53.22 | Helped |
| Center the network by subtracting its initial predictions (ablation) | CNN-GAP · CIFAR-10 | accuracy: 80.26 -> 81.25 | No clear effect |
| Use a large learning rate (ablation) | Finite-width networks · CIFAR-10 | accuracy: +1% to +5% | Mixed |
| Add L2 weight decay (ablation) | Finite-width networks · CIFAR-10 | accuracy: +1%-2% | Helped |
| Combine L2 regularization with early stopping (ablation) | CNN-VEC finite-width networks · CIFAR-10 | accuracy: 10% - 15% | Helped |
| Use L2 regularization with NTK parameterization instead of standard parameterization (ablation) | Finite-width networks · CIFAR-10 | accuracy: by up to 6% | Helped |
| Scale standard-parameterization L2 layerwise by layer width (ablation) | Finite-width standard-parameterization networks · CIFAR-10 | accuracy | Helped |
| Increase width (ablation) | CNN-VEC with NTK parameterization · CIFAR-10 | accuracy | Mixed |
Closest papers by meaning
- Towards a General Theory of Infinite-Width Limits of Neural Classifiers (ICML 2020 · similarity 0.91)
- Dynamics of Finite Width Kernel and Prediction Fluctuations in Mean Field Neural Networks (NeurIPS 2023 · similarity 0.91)
- An Infinite-Width Analysis on the Jacobian-Regularised Training of a Neural Network (ICML 2024 · similarity 0.90)
- Adaptive kernel predictors from feature-learning infinite limits of neural networks (ICML 2025 · similarity 0.89)
- Critical feature learning in deep neural networks (ICML 2024 · similarity 0.88)
- On the Surprising Effectiveness of Large Learning Rates under Standard Width Scaling (NeurIPS 2025 · similarity 0.88)
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.