When it comes to Huawei's "genius boy", the topic can be said to be full. But behind the million-level annual salary, what kind of work the "genius teenagers" do on a daily basis has always remained mysterious.
Now, the following is finally here - for the first time, Huawei has taken the initiative to disclose the latest trends of "genius boy": Zhong Zhao, who joined the company in 2019 and received a 2 million offer, led the team to apply AutoML algorithm research in less than a year. Thousands of Huawei Mate series and P series mobile phones.
This means that Zhong Zhao's team has successfully opened up the precedent for large-scale commercial use of AutoML.
As soon as the news came out, there must be another wave of topic explosions.
After less than a year of employment, the algorithm is used for tens of millions of Huawei mobile phones
The specific value method starts with Zhong Zhao's own research work.
In fact, Zhong Zhao solved a major pain point in image pixel processing algorithms—the balance between algorithm accuracy and model size.
If it is solved, pixel processing algorithms can be deployed to mobile phones to accelerate the speed of image processing such as spatial enhancement and super-resolution.
However, unlike common CV algorithms such as object detection and image classification, the production of such models requires a deep understanding of pixel-related properties.
In the direction of AutoML, there are already many applications of CV algorithms such as image classification and target recognition. However, specific to the pixel algorithm, no team has successfully applied AutoML to large-scale applications.
Pixels include a very large number of properties, such as color, brightness, and so on. Algorithms process pixels, it can be said that the most basic elements of the image need to be processed.
Therefore, this type of algorithm requires very high precision, and many Huawei experts have not been able to successfully overcome it before.
Zhong Zhao led the team and successfully applied AutoML technology to image pixel processing algorithms.
What is AutoML?
AutoML (Automated Machine Learning), which simply means "designing AI with AI", has become a hot research topic since 2014. In 2018, the technology gradually entered the stage of commercial trial acceleration.
In fact, before Zhong Zhao came to Huawei, Huawei's Noah's Ark Lab was already conducting related research in the direction of AutoML.
The laboratory has developed a full-process AutoML algorithm set VEGA, among which algorithms such as "Efficient Classification Network Search Scheme Based on Hardware Constraints (CARS)" and "Lightweight Super-Separation Network Structure Search (ESR-EA)" belong to NAS category.
And AutoML is exactly the direction of Zhong Zhao's research during his Ph.D.
In 2019, Zhong Zhao and Huawei, who were also working on AutoML at the time, "hit it right away". With the accumulation of his internship at SenseTime, he joined Huawei as a "genius boy" with an annual salary of 2.01 million and served as the leader of the AutoML research group. Inside, it breaks through the difficulty of this pixel processing algorithm.
Subsequently, Zhong Zhao led the team to develop an end-to-end pixel-level AutoML pipeline within two years of joining the company.
According to Huawei, this technology can "reduce the complexity of the video photography prototype algorithm by a hundred times when the academic and industry can only achieve 2-3 times." Will be used for more products.
Not only this research, Zhong Zhao has also made a lot of achievements in the mobile visual model.
Historically, there have been two main approaches to designing visual models for mobile:
One is to manually design lightweight network structures, such as ShuffleNet, MobileNetV3, etc., which have made some progress.
However, Zhong Zhao's team found that there is still redundancy between the convolution kernels of these models, which limits the speed of the model.
The other method is to compress the model, and obtain a small model with a similar structure to the large model through pruning, distillation and other means.
However, this method will reduce the accuracy, and it is difficult to meet the requirements of high-end mobile phones.
After Zhong Zhao came to Huawei, he led the team to propose a dynamic method for adaptively generating convolution kernels based on the content of images.
This method can significantly reduce the amount of computation while maintaining accuracy, ranging from 37% to 71.3% for different CNN networks.
In addition, in terms of data augmentation, Zhong Zhao also studied an adversarial automatic data augmentation method at Huawei, which was published on ICLR 2020.
Now, behind the success of these studies, it is also inseparable from Zhong Zhao's own efforts.
Family learning origin, learning computer since childhood
Zhong Zhao was born in 1991 into a family deeply influenced by computer science.
He majored in software engineering at Huazhong University of Science and Technology as an undergraduate, and won the first prize in the National Mathematical Contest in Modeling for Undergraduates in his junior year.
Zhong Zhao's father, a computer scientist, was a student of Qian Sanqiang and He Zehui.
Zhong Zhao developed a strong interest under the training of his father, and began to learn some programming knowledge. Growing up in this environment, it is no surprise that he chose a computer-related major in college.
During his undergraduate period, he also teamed up with his classmates to do some programming projects, such as the on-campus version of the drift bottle developed based on WeChat, which was very popular among the classmates.
After graduation, he came to the Institute of Automation, Chinese Academy of Sciences, under the tutelage of Deputy Director Liu Chenglin.
In 2018, his paper during his internship at SenseTime was selected for CVPR Oral and gave a keynote report at the conference. That year, only single-digit papers were selected for Oral in China.
In this paper, he proposed a block generation method for automatically building high-performance neural networks, which has been cited more than 400 times.
This is also the first paper published by Zhong Zhao in the relatively emerging direction of AutoML.
Later, his research direction gradually focused on this, and his doctoral dissertation was also titled "Deep Neural Network Architecture: From Manual Design to Automatic Learning".
Up to now, he has published many papers related to AutoML in international journal conferences such as IEEE T PATTERN ANAL, ICLR, iCCV, and NeurIPS.
In fact, there has been no shortage of external voices questioning the qualifications of the "genius boy" before.
There are anonymous users on Zhihu, who once questioned the actual ability of Zhong Zhao and other "genius boys", thinking that Huawei is "buying horse bones with money":
An IT industry insider who used to work with Huawei's "genius boy" in the same laboratory said.
Huawei's annual salary is indeed ridiculously high. In our laboratory, the average annual salary for doctoral graduates is about 600,000-800,000 yuan, and for master's degree is about 400,000 yuan.
But in the computer industry, they are under a lot of pressure even if they spend as much money as they can do.
This time, the official took the initiative to disclose the latest research results of the "genius boy" Zhong Zhao, and apply the relevant results to Huawei products, which is not only an affirmation of Zhong Zhao himself, but also shows Huawei's own confidence in this plan.
Zhong Zhao's "genius plan" is just a microcosm of Huawei's strengthening of R&D layout.
"These teenagers are like 'loach', revitalizing our organization and activating our team." Ren Zhengfei once said that in the next 3 to 5 years, I believe that Huawei will be completely new, all "change guns and guns", and we must win this "war".
Huawei once publicly explained how to cultivate "genius boys" in a video. Tian Qi, chief scientist in the field of artificial intelligence of Huawei Cloud, said in the video that "genius teenagers" mostly refer to doctoral students who have just graduated from the age of 25 to 30. At this time, they have the best physical strength, intelligence, and innovation ability.
In the process of training, the first step is to communicate and understand their strengths; the second step, "use good steel on the cutting edge", introduce business pain points and difficulties to them, and do a good job in connecting innovation from 0 to 1 to N ; Third, encourage them to actively discover and solve problems; fourth, provide a good laboratory atmosphere and encourage free thinking and discussion. Huawei's expectation for talented young people is to "do the strongest contact research, implement the results of these basic researches into the industry, and finally deposit them on the AI platform".
According to Huawei's official website, in addition to Zhong Zhao, Zuo Pengfei and Li Yi, who were selected as Huawei's "Genius Boys" in 2019, are currently engaged in cloud storage research and operating system formal verification at Huawei.
In addition to talented young people, Huawei has also invested in the cultivation of talents such as scientists for many years. According to public data, Huawei has at least more than 700 mathematicians, more than 800 physicists, more than 120 chemists, 15,000 people engaged in basic research, and more than 60,000 product R&D personnel. At the same time, Huawei has also cooperated with more than 300 universities, more than 900 research institutions and companies around the world, implemented 7,840 projects, invested 1.8 billion US dollars, and signed more than 1,000 R&D cooperation contracts for external payments.
After this, will more research results of "genius teenagers" be made public?
We will wait and see.
——————
Original: Lao Wang丨[Public Account: Hongmeng Developer Laowang] Huawei Certified Instructor / Tencent Certified Instructor / Hongmeng Development Pioneer