Related Experiment Video
Updated: May 6, 2026
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
11.0K
Automated deep learning by recurrent hyperparameter optimization
Zhanzhan Cheng1,2, Yuyi Cheng1,3, Chenbo Zhang3
1EZVIZ, Hangzhou, China.
Nature Communications
|May 4, 2026
Summary
Rocket automates deep learning hyperparameter optimization (HPO) using reinforcement learning, eliminating the need for domain expertise. This novel framework achieves state-of-the-art results efficiently, significantly reducing time and cost in industrial applications.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Deep Learning
Background:
- Optimizing deep learning hyperparameters is complex, demanding significant expertise and resources.
- Current hyperparameter optimization (HPO) methods have limitations in handling mixed-type hyperparameters, scalability, and automation.
Purpose of the Study:
- To introduce Rocket, a novel recurrent HPO framework for automated tuning of mixed-type hyperparameters.
- To enable deep learning models to achieve state-of-the-art performance without prior domain knowledge.
Main Methods:
- Utilizes self-play reinforcement learning with a policy agent learning from historical interactions.
- Implements a reward approximation mechanism using data subsets to accelerate policy learning on large datasets (up to 80X speedup).
Main Results:
- Rocket achieved state-of-the-art performance across 8 deep learning tasks and 32 benchmarks, matching expert-tuned models.
- Demonstrated significant efficiency gains in industrial deployment, reducing optimization time by 13.4-fold and cost by 73%.
Conclusions:
- Rocket effectively automates hyperparameter optimization for deep learning models.
- The framework offers a scalable, efficient, and domain-agnostic solution for HPO, outperforming existing methods.