Research map / Optimization / Implicit bias and generalization
Chaotic Regularization and Heavy-Tailed Limits for Deterministic Gradient Descent
Soon Hoe Lim, Yijun Wan, Umut Simsekli
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 MPGD instead of full-batch GD | ResNet-18 · CIFAR-10 validation | validation accuracy: 73.10 -> 75.97 | Improved |
| Use MPGD instead of full-batch GD | ResNet-18 · CIFAR-10 | accuracy gap: 8.99 -> 5.64 | Improved |
| Use MPGD instead of vanilla GD | Shallow neural net, width 16 · Airfoil Self-Noise test | test RMSE: 0.4309 -> 0.3810 | Improved |
| Use MPGD instead of vanilla GD | Shallow neural net, width 16 · Airfoil Self-Noise | RMSE gap: 0.2411 -> 0.2256 | Improved |
| Add uncorrelated Gaussian perturbations instead of MPGD | Full-batch GD with ResNet-18 · CIFAR-10 validation | validation accuracy: 73.10 -> 70.58 | Hurt |
| Set γ = 0.7 rather than γ = 0.6 (ablation) | MPGD with shallow neural net · Airfoil Self-Noise test | test RMSE: 0.3810 -> 0.4600 | Hurt |
| Set γ = 0.55 rather than γ = 0.6 (ablation) | MPGD with ResNet-18 · CIFAR-10 validation | validation accuracy: 75.78 -> 75.97 | No clear effect |
| Increase γ from 0.6 to 0.7 (ablation) | MPGD · Airfoil Self-Noise and CIFAR-10 | test performance: 0.3810 -> 0.4600; 75.78 -> 73.72 | Mixed |
| Replace MPGD perturbations with Gaussian perturbations | Full-batch GD · Widening valley loss | optimization success | Hurt |
| Use MPGD perturbations rather than Gaussian perturbations | Full-batch GD · CIFAR-10 validation | validation accuracy: 70.58 -> 75.97 | Improved |
Closest papers by meaning
- Algorithmic Stability of Heavy-Tailed SGD with General Loss Functions (ICML 2023 · similarity 0.90)
- Emergence of heavy tails in homogenized stochastic gradient descent (NeurIPS 2024 · similarity 0.90)
- Hausdorff Dimension, Heavy Tails, and Generalization in Neural Networks (NeurIPS 2020 · similarity 0.90)
- Error Analysis Affected by Heavy-Tailed Gradients for Non-Convex Pairwise Stochastic Gradient Descent (AAAI 2025 · similarity 0.89)
- The Global Convergence Time of Stochastic Gradient Descent in Non-Convex Landscapes: Sharp Estimates via Large Deviations (ICML 2025 · similarity 0.88)
- Long-time asymptotics of noisy SVGD outside the population limit (ICLR 2025 · similarity 0.87)
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.
