Research map / Optimization / Optimizers
Accelerated Training via Incrementally Growing Neural Networks using Variance Transfer and Learning Rate Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Grow the network with variance transfer and stage-wise learning-rate adaptation | ResNet-20 · CIFAR-10 | test accuracy (%): 92.62\pm0.15 -> 92.53\pm 0.11 | No clear effect |
| Grow the network with variance transfer and stage-wise learning-rate adaptation | VGG-11 · CIFAR-10 | test accuracy (%): 92.14\pm0.22 -> 92.34\pm0.15 | Helped |
| Grow the network with variance transfer and stage-wise learning-rate adaptation | ResNet-18 · CIFAR-100 | test accuracy (%): 78.36\pm0.12 -> 78.12\pm 0.15 | No clear effect |
| Grow the network with variance transfer and stage-wise learning-rate adaptation | VGG-19 · CIFAR-100 | test accuracy (%): 72.59\pm 0.23 -> 73.26\pm0.14 | Helped |
| Grow the network with variance transfer and stage-wise learning-rate adaptation | ResNet-50 · ImageNet | train cost (%): 100 -> 60.12 | Helped |
| Grow the network and apply buffer preservation plus rate adaptation (ablation) | Transformer · IWSLT'14 De-En | BLEU: 31.62 \pm 0.16 -> 32.01 \pm 0.16 | Helped |
| Preserve optimizer buffers across growth phases (ablation) | Transformer · IWSLT'14 De-En | BLEU: 31.44 \pm 0.18 -> 31.62 \pm 0.16 | Helped |
| Use variance transfer instead of standard initialization (ablation) | ResNet-20 · CIFAR-10 | test accuracy (%): 91.62\pm 0.12 -> 92.00\pm 0.10 | Helped |
| Use feature-variance constrained weight rescaling instead of default variance transfer (ablation) | ResNet-20 · CIFAR-10 | test accuracy (%): 92.00\pm 0.10 -> 91.93 \pm 0.12 | Hurt |
| Apply stage-wise learning-rate adaptation (ablation) | ResNet-20 · CIFAR-10 | test accuracy (%): 91.62\pm 0.12 -> 92.53\pm 0.11 | Helped |
| Use layer-wise LARS adaptation without stage-wise adaptation (ablation) | ResNet-20 · CIFAR-10 | test accuracy (%): 91.95\pm 0.09 -> 91.32 \pm 0.11 | Hurt |
| Use max(1, norm of newly added weights) as the rate adaptation factor (ablation) | ResNet-20 · CIFAR-10 | test accuracy (%): 91.62\pm 0.12 -> 91.42 \pm 0.12 | Hurt |
Closest papers by meaning
- Weight Decay may matter more than µP for Learning Rate Transfer in Practice (ICLR 2026 · similarity 0.88)
- Deep Linear Network Training Dynamics from Random Initialization: Data, Width, Depth, and Hyperparameter Transfer (ICML 2025 · similarity 0.87)
- On the Weight Dynamics of Deep Normalized Networks (ICML 2024 · similarity 0.86)
- MGUP: A Momentum-Gradient Alignment Update Policy for Stochastic Optimization (NeurIPS 2025 · similarity 0.86)
- MARS: Unleashing the Power of Variance Reduction for Training Large Models (ICML 2025 · similarity 0.85)
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 2025 · similarity 0.85)
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.