Research map / Optimization / Implicit bias and generalization
Analyzing Generalization of Neural Networks through Loss Path Kernels
Yilan Chen, Wei Huang, Hao Wang, Charlotte Loh, Akash Srivastava, Lam M. Nguyen, Tsui-Wei Weng
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 |
|---|---|---|---|
| Use Gene(w, S) after 1 epoch instead of TENAS to pick architectures | random search NAS · NAS-Bench-201 CIFAR-10 | test accuracy: 93.08±0.15 -> 93.68±0.12 | Improved |
| Use Gene(w, S) after 2 epochs instead of TENAS | random search NAS · NAS-Bench-201 CIFAR-10 | test accuracy: 93.08±0.15 -> 93.79±0.18 | Improved |
| Use Gene(w, S) after 1 epoch instead of TENAS | random search NAS · NAS-Bench-201 CIFAR-100 | test accuracy: 70.37±2.40 -> 72.02±1.43 | Improved |
| Use Gene(w, S) after 2 epochs instead of TENAS | random search NAS · NAS-Bench-201 CIFAR-100 | test accuracy: 70.37±2.40 -> 72.76±0.33 | Improved |
| Use Gene(w, S) after 1 epoch instead of RS + LGA3 | random search NAS · NAS-Bench-201 CIFAR-100 | test accuracy: 69.77 -> 72.02±1.43 | Improved |
| Use Gene(w, S) after 2 epochs (vs 1 epoch) (ablation) | RS + Gene(w, S) · NAS-Bench-201 CIFAR-10 | test accuracy: 93.68±0.12 -> 93.79±0.18 | Improved |
| Use Gene(w, S) after 2 epochs (vs 1 epoch) (ablation) | RS + Gene(w, S) · NAS-Bench-201 CIFAR-100 | best-run accuracy: 73.15 -> 73.15 | No clear effect |
| Compare Gene(w, S) search with training each architecture to convergence | RS + Gene(w, S) · NAS-Bench-201 CIFAR-10 | GPU hours per architecture: 1.83 -> 0.036 | Improved |
| Compare Gene(w, S) search with training each architecture to convergence | RS + Gene(w, S) · NAS-Bench-201 CIFAR-100 | GPU hours per architecture: 2.56 -> 0.037 | Improved |
| Replace norm-based bound with loss path kernel bound | generalization bound · MNIST 1 vs 7, two-layer NN | bound value: 140.7 -> 0.03 | Improved |
| Replace VC dimension bound with loss path kernel bound | generalization bound · MNIST 1 vs 7, two-layer NN | bound value: 55957.3 -> 0.03 | Improved |
| Replace Cao & Gu NTK-based bound with loss path kernel bound | generalization bound · MNIST 1 vs 7, 1000-hidden NN | bound value: 1.44 -> 0.026 | Improved |
Closest papers by meaning
- Generalization Bound of Gradient Flow through Training Trajectory and Data-dependent Kernel (NeurIPS 2025 · similarity 0.91)
- Disentangling Trainability and Generalization in Deep Neural Networks (ICML 2020 · similarity 0.89)
- Understanding Generalization and Optimization Performance of Deep CNNs (ICML 2018 · similarity 0.88)
- How much does Initialization Affect Generalization? (ICML 2023 · similarity 0.88)
- Do Neural Networks Need Gradient Descent to Generalize? A Theoretical Study (NeurIPS 2025 · similarity 0.87)
- Deeper or Wider: A Perspective from Optimal Generalization Error with Sobolev Loss (ICML 2024 · similarity 0.86)
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.
