Research map / Optimization / Implicit bias and generalization
Understanding Multi-phase Optimization Dynamics and Rich Nonlinear Behaviors of ReLU Networks
Mingze Wang, Chao Ma
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 |
|---|---|---|---|
| Shrink angle Delta between classes (4pi/45 to (3/4)^3 * 4pi/45) (ablation) | Two-layer ReLU net, GF · Plateau duration T_plat, p=4 | T_plat: 1.96\times10^4 -> 12.87\times10^4 | Hurt |
| Shrink angle Delta (ablation) | Two-layer ReLU net, GF · Time to convergence T_III, p=4 | T_III: 4.98\times10^4 -> 15.63\times10^4 | Hurt |
| Increase dominance ratio p from 6 to 12 (ablation) | Two-layer ReLU net, GF · Plateau duration, Delta=pi/15 | T_plat: 6.14\times10^4 -> 17.61\times10^4 | Hurt |
| Increase p from 6 to 12 (ablation) | Two-layer ReLU net, GF · Time T_III, Delta=pi/15 | T_III: 8.92\times10^4 -> 27.68\times10^4 | Hurt |
| Fitted theoretical estimate vs realistic T_plat over Delta | Theorem bound Theta(1/Delta^2) · p=4 | T_plat: 12.87\times10^4 vs 12.89\times10^4 | Improved |
| Add small stochastic angle noise Unif([0,Delta/4]) to data (ablation) | Two-layer ReLU net, GF · Noisy dataset, accuracy | training accuracy plateaus: 12/15, 13/15, 14/15 | Mixed |
Closest papers by meaning
- Gradient Descent on Two ReLU Neurons: Global Landscape and Bifurcation Dynamics (NeurIPS 2026 · similarity 0.91)
- Gradient flow dynamics of shallow ReLU networks for square loss and orthogonal inputs (NeurIPS 2022 · similarity 0.90)
- Neural Characteristic Activation Analysis and Geometric Parameterization for ReLU Networks (NeurIPS 2024 · similarity 0.90)
- Simplicity Bias and Optimization Threshold in Two-Layer ReLU Networks (ICML 2025 · similarity 0.89)
- Convex Relaxations of ReLU Neural Networks Approximate Global Optima in Polynomial Time (ICML 2024 · similarity 0.88)
- Expressivity of ReLU-Networks under Convex Relaxations (ICLR 2024 · 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.
