Research map / Optimization / Implicit bias and generalization
On the Generalization of Stochastic Gradient Descent with Momentum
Ali Ramezani-Kebrya, Kimon Antonakopoulos, Volkan Cevher, Ashish Khisti, Ben Liang
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 |
|---|---|---|---|
| add momentum only for the early phase of training (SGDEM) instead of no momentum (ablation) | SGD · notMNIST (tuned, td=3000) | test error: 0.3596 ± 0.0142 -> 0.2892 ± 0.0042 | Improved |
| restrict momentum to the first td iterations (SGDEM) instead of the entire training (SGDM) (ablation) | SGDM · notMNIST (tuned, td=3000) | test error: 0.3194 ± 0.0030 -> 0.2892 ± 0.0042 | Improved |
| add momentum for the entire training (SGDM) instead of early only, for large momentum 0.99 on nonconvex loss (ablation) | SGDEM · notMNIST nonconvex | test error | Hurt |
| increase momentum parameter mu in strongly convex logistic regression (ablation) | SGD · notMNIST logistic regression (generalization error) | generalization error | Hurt |
| add momentum (mu 0.5 or 0.9) to SGD in strongly convex logistic regression (ablation) | SGD · notMNIST logistic regression (training error) | training error | Improved |
| add momentum to SGD for many epochs in logistic regression (ablation) | SGD · notMNIST logistic regression (n=500) | test accuracy | No clear effect |
| apply momentum 0.9 for the first 50 epochs only instead of all 90 epochs (ablation) | SGDM · ImageNet ResNet-18 | generalization error | Improved |
| choose intermediate td for early momentum, small mu_d (ablation) | SGDEM · CIFAR10 ResNet-20 | test error | Improved |
Closest papers by meaning
- An Improved Analysis of Stochastic Gradient Descent with Momentum (NeurIPS 2020 · similarity 0.94)
- The Marginal Value of Momentum for Small Learning Rate SGD (ICLR 2024 · similarity 0.92)
- Random Scaling and Momentum for Non-smooth Non-convex Optimization (ICML 2024 · similarity 0.91)
- Gradient correlation is a key ingredient to accelerate SGD with momentum (ICLR 2025 · similarity 0.91)
- Accelerated Convergence of Stochastic Heavy Ball Method under Anisotropic Gradient Noise (ICLR 2024 · similarity 0.91)
- Perfect Parallelization in Mini-Batch SGD with Classical Momentum Acceleration (NeurIPS 2026 · similarity 0.89)
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.
