Research map / Optimization / Implicit bias and generalization
Imitating Deep Learning Dynamics via Locally Elastic Stochastic Differential Equations
Jiayao Zhang, Hua Wang, Weijie J Su
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 logits-as-features (L-model) instead of isotropic features (I-model) to estimate dynamics | LE-ODE simulations on GeoMNIST and CIFAR · GeoMNIST and CIFAR, K=3 | logit-dynamics fidelity | Improved |
| Use isotropic-feature-based estimates instead of logits-as-features estimates (ablation) | LE-ODE simulations on GeoMNIST and CIFAR · GeoMNIST and CIFAR, K=3 | logit-dynamics fidelity | Hurt |
| Increase label corruption toward random labeling | AlexNet training · GeoMNIST | validation accuracy: 2/3 | Hurt |
| Increase label corruption toward random labeling | AlexNet training · GeoMNIST | feature separability: around p = 0.6 | Hurt |
| Make intra-class elasticity greater than inter-class elasticity | LE-SDE · Binary classification | asymptotic feature separability: α > β | Improved |
| Set intra-class elasticity equal to or below inter-class elasticity | LE-SDE · Binary classification | asymptotic feature separability: α ≤ β | Hurt |
Closest papers by meaning
- Neural Mechanics: Symmetry and Broken Conservation Laws in Deep Learning Dynamics (ICLR 2021 · similarity 0.90)
- Feature Learning Dynamics in Infinite-Depth Neural Networks (NeurIPS 2026 · similarity 0.88)
- Three Mechanisms of Feature Learning in a Linear Network (ICLR 2025 · similarity 0.87)
- Neural Thermodynamics: Entropic Forces in Deep and Universal Representation Learning (NeurIPS 2025 · similarity 0.87)
- Global Convergence and Rich Feature Learning in $L$-Layer Infinite-Width Neural Networks under $\mu$ Parametrization (ICML 2025 · similarity 0.86)
- Analyzing Neural Scaling Laws in Two-Layer Networks with Power-Law Data Spectra (ICLR 2025 · 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.
