AI Paper Atlas

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

AAAI 2018 · Architectures and efficiency › RNNs and normalization · approach: Efficient network compression and training · 122 citations (OpenAlex)

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 changedWhereMeasuredEffect
Train a light net jointly with a booster and guide it with booster logitsWRN-16-1(b) · CIFAR-10test error: 8.77 -> 7.87Improved
Train the light net jointly with the booster (ablation)Rocket · CIFAR-10test error: 8.04 -> 7.87Improved
Share lower-layer parameters between light net and booster (ablation)Rocket · CIFAR-10test error: 8.06 -> 7.87Improved
Add gradient block to prevent hint-loss gradients from updating booster-specific parameters (ablation)Rocket · CIFAR-10test error: 8.50 -> 7.87Improved
Use logits-MSE rather than probability-MSE as the hint loss (ablation)Rocket · CIFAR-10test error: 8.32 -> 7.87Improved
Use logits-MSE rather than KD as the hint loss (ablation)Rocket · CIFAR-10test error: 7.98 -> 7.87Improved
Add KD guidance from a pretrained booster to RocketWRN-16-1(b) Rocket · CIFAR-10test error: 7.87 -> 7.52Improved
Add KD guidance to RocketWRN-16-1(b) Rocket · SVHNtest error: 2.29 -> 2.20Improved
Add KD guidance to RocketWRN-16-1(b) Rocket · CIFAR-100test error: 33.3 -> 33.0Improved
Use Rocket training for the light networkDIN-style light network · Advertisement datasetGAUC: 0.632 -> 0.635Improved
Use the booster-only network instead of the Rocket light networkDIN-style network · Advertisement datasetinference time of FC Layers: 7.6 ms -> 23.2 msHurt
Use Rocket training instead of training the light and booster networks separatelyCIFAR-10 training · CIFAR-10training time: 7275.3s -> 6153.0sImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy