How Huawei’s “Genius Teen” Scaled AutoML to Millions of Phones

AutoML on Millions of Huawei Phones

In 2019, Zhong Zhao joined Huawei with a 2‑million‑yuan offer and, in less than a year, led his team to apply AutoML algorithms to tens of millions of Huawei Mate and P series phones.

Huawei announced that this achievement marked the first large‑scale commercial use of AutoML.

The breakthrough addressed a major pain point in image‑pixel processing: balancing algorithm accuracy with model size, enabling deployment of pixel‑level processing for tasks such as spatial enhancement and super‑resolution on mobile devices.

Unlike typical computer‑vision models like object detection or image classification, pixel‑level models require deep understanding of pixel attributes (color, brightness, etc.), making high‑precision algorithms difficult to achieve.

Zhong’s team successfully applied AutoML to these pixel‑processing algorithms, a domain where prior attempts had failed.

AutoML (Automated Machine Learning) is essentially “AI designing AI”. It became a hot research topic in 2014 and entered the trial‑commercial acceleration phase in 2018.

Before Zhong, Huawei Noah’s Ark Lab had already been researching AutoML, developing a full‑process AutoML suite called VEGA, which includes hardware‑constrained efficient classification network search (CARS) and lightweight super‑resolution network search (ESR‑EA), both falling under Neural Architecture Search (NAS).

Zhong’s Ph.D. research focused on AutoML, and after joining Huawei as the AutoML research group leader, he broke the pixel‑processing barrier within a year.

Within two years, his team built an end‑to‑end pixel‑level AutoML pipeline that can reduce the complexity of video‑photography prototype algorithms by a factor of 100, far surpassing the 2‑3× improvements typical in academia and industry.

Beyond this, Zhong contributed to mobile visual models. Traditional lightweight networks (e.g., ShuffleNet, MobileNetV3) still contain redundant convolution kernels, limiting speed, while model‑compression techniques (pruning, distillation) often sacrifice accuracy.

His team proposed a dynamic method that generates convolution kernels adaptively based on image content, cutting computation by 37%‑71.3% across various CNNs without losing precision.

He also introduced an adversarial automatic data‑augmentation technique, published at ICLR 2020.

Early Life and Education

Zhong was born in 1991 into a family deeply rooted in computer science; his father was a computer scientist and a student of Qian Sanqiang and He Zehui.

He studied Software Engineering at Huazhong University of Science and Technology, winning a first‑prize award in the Hubei division of the National College Student Mathematical Modeling Competition.

After graduation, he joined the Institute of Automation, Chinese Academy of Sciences, under Deputy Director Liu Chenglin.

During an internship at SenseTime in 2018, his first‑author paper was selected for a CVPR Oral presentation, a rare honor for Chinese researchers.

His early work introduced a block‑generation method for automatically constructing high‑performance neural networks, which has been cited over 400 times.

His Ph.D. dissertation, “Deep Neural Network Structures: From Manual Design to Automated Learning,” and subsequent publications in IEEE Transactions on Pattern Analysis, ICLR, ICCV, and NeurIPS solidify his standing in the AutoML community.