Research map / Optimization / Implicit bias and generalization
Iterate Averaging in the Quest for Best Test Error
Diego Granziol, Nicholas P. Baskerville, Xingchen Wan, Samuel Albanie, Stephen Roberts
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 iterate averaging (GadamX vs AdamW) with partially decoupled weight decay | AdamW · CIFAR-100, ResNeXt-29 | test accuracy: 80.16 -> 83.27 | Improved |
| add iterate averaging and strong weight decay to Adam (Gadam vs AdamW) | AdamW · ImageNet, ResNet-50 | top-1 accuracy: 74.04 -> 76.79 | Improved |
| add iterate averaging to SGD (SWA vs SGD) | SGD · ImageNet 32x32, WRN-28-10 | top-1 accuracy: 61.33 -> 62.32 | Improved |
| use GadamX instead of SWA | SWA · CIFAR-100, VGG-16 | test accuracy: 74.57 -> 76.85 | Improved |
| use GadamX instead of SWA | SWA · CIFAR-100, PRN-110 | test accuracy: 77.92 -> 77.90 | No clear effect |
| use plain Gadam instead of tuned SGD | SGD · ImageNet 32x32, WRN-28-10 | top-1 accuracy: 61.33 -> 60.50 | Hurt |
| tune the averaging start T_avg (ablation) | GadamX · CIFAR-100, PRN-110 | test accuracy: 77.90 -> 79.41 | Improved |
| tune the averaging start T_avg (ablation) | GadamX · CIFAR-100, VGG-16 | test accuracy: 77.90 -> 77.22 | Improved |
| reduce numerical stability constant to 1e-4 in GadamX (partial adaptivity) (ablation) | GadamX · ImageNet, ResNet-50 | top-1 accuracy: 77.19 -> 77.31 | Improved |
| raise base learning rate to 0.003 in SGD-like Gadam (ablation) | Gadam · ImageNet, ResNet-50 | top-1 accuracy: 75.52 -> 76.53 | Improved |
| average once per iteration instead of less often (ablation) | Gadam · PTB LSTM validation perplexity | perplexity | No clear effect |
| use automatic averaging trigger and early stopping (Auto) with Adam base vs SGD base (ablation) | Gadam-Auto vs SWA-Auto · CIFAR-100, VGG-16 | test accuracy: 73.93 -> 75.39 | Improved |
Closest papers by meaning
- Estimating Generalization Performance Along the Trajectory of Proximal SGD in Robust Regression (NeurIPS 2024 · similarity 0.84)
- When, Where and Why to Average Weights? (ICML 2025 · similarity 0.84)
- On the Last-Iterate Convergence of Shuffling Gradient Methods (ICML 2024 · similarity 0.84)
- Generalization Analysis of Stochastic Weight Averaging with General Sampling (ICML 2024 · similarity 0.83)
- Revisiting the Last-Iterate Convergence of Stochastic Gradient Methods (ICLR 2024 · similarity 0.83)
- A General-Purpose Theorem for High-Probability Bounds of Stochastic Approximation with Polyak Averaging (NeurIPS 2025 · similarity 0.83)
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.
