Research map / Optimization / Nonconvex and smooth optimization
Stagewise Training Accelerates Convergence of Testing Error Over SGD
Zhuoning Yuan, Yan Yan, Rong Jin, Tianbao Yang
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 stagewise training instead of SGD with a polynomially decaying step size | Stagewise training on CIFAR-10 and CIFAR-100 with ResNets and ConvNets · CIFAR-10 and CIFAR-100 | testing error | Improved |
| Use stagewise training instead of SGD with a polynomially decaying step size | Stagewise training with convex losses · real-sim and E2006-tfidf | testing error | Improved |
| Add explicit regularization at each stage (ablation) | Stagewise SGD (V1) · CIFAR-10 and CIFAR-100 | generalization error | Improved |
| Add explicit regularization at each stage (ablation) | Stagewise SGD (V1) · CIFAR-10 and CIFAR-100 | training error | Hurt |
| Return the averaged solution at each stage instead of the last solution (ablation) | Stagewise SGD (V2) · CIFAR-10 and CIFAR-100 | generalization error | Improved |
| Return the averaged solution at each stage instead of the last solution (ablation) | Stagewise SGD (V2) · CIFAR-10 and CIFAR-100 | training error | Hurt |
| Use explicit regularization at each stage and return the last solution (ablation) | Stagewise SGD (V1) · CIFAR-10 and CIFAR-100 | testing error | Improved |
| Use an averaged solution at each stage rather than the last solution (ablation) | START on convex losses · real-sim and E2006-tfidf | testing error | No clear effect |
Closest papers by meaning
- Universal Stagewise Learning for Non-Convex Problems with Convergence on Averaged Solutions (ICLR 2019 · similarity 0.90)
- Special Properties of Gradient Descent with Large Learning Rates (ICML 2023 · similarity 0.88)
- SGD Converges to Global Minimum in Deep Learning via Star-convex Path (ICLR 2019 · similarity 0.88)
- An Analysis of Constant Step Size SGD in the Non-convex Regime: Asymptotic Normality and Bias (NeurIPS 2021 · similarity 0.88)
- Fast Last-Iterate Convergence of SGD in the Smooth Interpolation Regime (NeurIPS 2025 · similarity 0.87)
- Rapid Overfitting of Multi-Pass SGD in Stochastic Convex Optimization (ICML 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.
