Research map / Architectures and efficiency / RNNs and normalization
Rocket Launching: A Universal and Efficient Framework for Training Well-Performing Light Net
Guorui Zhou, Ying Fan, Runpeng Cui, Weijie Bian, Xiaoqiang Zhu, Kun Gai
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 |
|---|---|---|---|
| Train a light net jointly with a booster and guide it with booster logits | WRN-16-1(b) · CIFAR-10 | test error: 8.77 -> 7.87 | Improved |
| Train the light net jointly with the booster (ablation) | Rocket · CIFAR-10 | test error: 8.04 -> 7.87 | Improved |
| Share lower-layer parameters between light net and booster (ablation) | Rocket · CIFAR-10 | test error: 8.06 -> 7.87 | Improved |
| Add gradient block to prevent hint-loss gradients from updating booster-specific parameters (ablation) | Rocket · CIFAR-10 | test error: 8.50 -> 7.87 | Improved |
| Use logits-MSE rather than probability-MSE as the hint loss (ablation) | Rocket · CIFAR-10 | test error: 8.32 -> 7.87 | Improved |
| Use logits-MSE rather than KD as the hint loss (ablation) | Rocket · CIFAR-10 | test error: 7.98 -> 7.87 | Improved |
| Add KD guidance from a pretrained booster to Rocket | WRN-16-1(b) Rocket · CIFAR-10 | test error: 7.87 -> 7.52 | Improved |
| Add KD guidance to Rocket | WRN-16-1(b) Rocket · SVHN | test error: 2.29 -> 2.20 | Improved |
| Add KD guidance to Rocket | WRN-16-1(b) Rocket · CIFAR-100 | test error: 33.3 -> 33.0 | Improved |
| Use Rocket training for the light network | DIN-style light network · Advertisement dataset | GAUC: 0.632 -> 0.635 | Improved |
| Use the booster-only network instead of the Rocket light network | DIN-style network · Advertisement dataset | inference time of FC Layers: 7.6 ms -> 23.2 ms | Hurt |
| Use Rocket training instead of training the light and booster networks separately | CIFAR-10 training · CIFAR-10 | training time: 7275.3s -> 6153.0s | Improved |
Closest papers by meaning
- A Physics-Inspired Optimizer: Velocity Regularized Adam (ICLR 2026 · similarity 0.80)
- Generalize for Future: Slow and Fast Trajectory Learning for CTR Prediction (AAAI 2024 · similarity 0.80)
- Accelerating neural network training: An analysis of the AlgoPerf competition (ICLR 2025 · similarity 0.80)
- Deep Learning Through A Telescoping Lens: A Simple Model Provides Empirical Insights On Grokking, Gradient Boosting & Beyond (NeurIPS 2024 · similarity 0.79)
- Fast and Slow Gradient Approximation for Binary Neural Network Optimization (AAAI 2025 · similarity 0.79)
- Near-optimal Sketchy Natural Gradients for Physics-Informed Neural Networks (ICML 2025 · similarity 0.79)
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.
