Research map / Optimization / Implicit bias and generalization
Three-stage Evolution and Fast Equilibrium for SGD with Non-degerate Critical Points
Yi Wang, Zhiren Wang
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 the normalized (spherical) SGD with constant effective learning rate instead of unnormalized dynamics | SGD with constant effective learning rate on sphere · Toy function (d=100) | time to equilibrium scaling in zeta | Improved |
| Equilibrium reached within O(ζ^-1) time rather than the O(λe^-1) bound | Underparametrized CNN on MNIST · MNIST d=332 | time to equilibrium | No clear effect |
| Overparametrized network with many basins: equilibria from different initial positions remain far apart | Overparametrized CNN on MNIST (d≈100k) · MNIST d≈100k | mean squared distance between equilibria from different initializations (V12) | Mixed |
Closest papers by meaning
- Stochastic Collapse: How Gradient Noise Attracts SGD Dynamics Towards Simpler Subnetworks (NeurIPS 2023 · similarity 0.88)
- From Gradient Flow on Population Loss to Learning with Stochastic Gradient Descent (NeurIPS 2022 · similarity 0.87)
- Emergence and scaling laws in SGD learning of shallow neural networks (NeurIPS 2025 · similarity 0.87)
- The Global Convergence Time of Stochastic Gradient Descent in Non-Convex Landscapes: Sharp Estimates via Large Deviations (ICML 2025 · similarity 0.86)
- What is the Long-Run Distribution of Stochastic Gradient Descent? A Large Deviations Analysis (ICML 2024 · similarity 0.86)
- High-dimensional Limit of SGD for Diagonal Linear Networks (NeurIPS 2026 · 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.
