Research map / Optimization / Optimizers
L4: Practical loss-based stepsize adaptation for deep learning
Michal Rolinek, Georg Martius
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 |
|---|---|---|---|
| Replace constant tuned stepsizes with default L4 adaptation for Adam or momentum SGD | Adam and momentum SGD · MNIST | test accuracy: 97.9 ± 0.3 -> 98.4 ± 0.1; 98.0 ± 0.4 -> 98.4 ± 0.1 | Mixed |
| Replace tuned constant stepsizes with default L4 adaptation | Adam and momentum SGD · Fashion MNIST | test accuracy: 93.1 ± 0.2 -> 93.35 ± 0.15; 93.0 ± 0.15 -> 93.25 ± 0.2 | Improved |
| Increase dropout from p=0.4 to p=0.7 (ablation) | L4 Adam and L4 Mom · Fashion MNIST | test accuracy: 93.35 ± 0.15 -> 93.6 ± 0.25; 93.25 ± 0.2 -> 93.45 ± 0.15 | Improved |
| Use smaller batch sizes instead of the main-paper batch size (ablation) | L4 Adam and L4 Mom · MNIST | training loss | Improved |
| Use smaller batch sizes (ablation) | L4 Adam and L4 Mom · DNC REPEAT-COPY | training loss | Improved |
| Use batch size 8 (ablation) | L4 Mom · DNC REPEAT-COPY | training loss | Mixed |
| Use L4-adapted stepsizes rather than best constant stepsizes | L4 Mom and L4 Adam · CIFAR-10 ResNet | test accuracy | Improved |
| Use default L4 adaptation instead of the repository's piecewise learning-rate policy | L4 Mom and L4 Adam · CIFAR-10 ResNet | test accuracy | Hurt |
| Use L4 adaptation rather than optimized constant learning rates | L4 Adam and L4 Mom · DNC REPEAT-COPY | training loss | Improved |
| Use L4 adaptation rather than optimized constant learning rates | L4 Adam and L4 Mom · Badly conditioned regression | time to convergence: 1.4 ± 0.5 vs 3.0 ± 1.4; 1.0 ± 0.02 vs 3.0 ± 1.4; 2.1± 0.4 vs 123 ± 64; 1.8 ± 0.2 vs 123 ± 64 | Mixed |
Closest papers by meaning
- Enhancing Optimizer Stability: Momentum Adaptation of The NGN Step-size (NeurIPS 2025 · similarity 0.87)
- Provable and Practical Online Learning Rate Adaptation with Hypergradient Descent (ICML 2025 · similarity 0.85)
- Stepsize anything: A unified learning rate schedule for budgeted-iteration training (NeurIPS 2025 · similarity 0.85)
- Adam Reduces a Unique Form of Sharpness: Theoretical Insights Near the Minimizer Manifold (NeurIPS 2025 · similarity 0.85)
- Understanding the Generalization of Stochastic Gradient Adam in Learning Neural Networks (NeurIPS 2025 · similarity 0.84)
- Prodigy: An Expeditiously Adaptive Parameter-Free Learner (ICML 2024 · similarity 0.84)
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.
