According to a recent report on artificial intelligence by McKinsey, in the coming decade, there will be tremendous opportunities for AI growth in new sectors in China. These are industries where innovation and R&D spending have traditionally lagged global counterparts, including automotive, transportation, logistics, manufacturing, enterprise software, healthcare, and life sciences.
As China looks to AI as a new engine of economic development, four “dragons” have emerged to lead the development of homegrown AI technologies—Cloudwalk, Yitu, SenseTime, and Megvii.
Among the quartet, the first to set up AI operations is Megvii, which was founded in Beijing in 2011 by three Tsinghua University graduates —Yin Qi, Tang Wenbin, and Yang Mu—who studied under computer scientist and computational theorist Andrew Chi-Chih Yao, who in 2000 won the Turing Award, the “Nobel Prize of computing.”
Known for its facial recognition technology, the Chinese developer of AI hardware and software has evolved to become a creator of intelligent ecosystems, optimizing the effectiveness of Internet-of-Things (IoT) devices in manufacturing, logistics, and urban management. Headquartered in Beijing with over 2,000 employees, Megvii is backed by e-commerce giant Alibaba, and maintains four research and development centers in China.
According to Yin Qi, co-founder and CEO of Megvii, the AI firm has three business verticals: personal IoT, city IoT which applies AI to facilitate the efficient management of city-wide operations, and supply chain IoT for operational efficiency in factory operations.
In 2020, the tech firm launched Brain++, an AI productivity platform that helps businesses build their own AI capabilities, deploy customized algorithms at scale, and accelerate their digital transformation.
Working as a foundational structure to provide support for Megvii’s algorithm training and model improvement processes, the deep learning framework is designed to solve pain points that businesses encounter when trying to digitize their operations at scale. The algorithm development platform covers the entire lifecycle of algorithm production, from data management to model optimization and scheduling.
Developing AI algorithms can be an arduous task, one that needs substantial efforts in research and systems engineering. Building an in-house framework from scratch also requires considerable time and organizational resources to ensure that the AI system runs smoothly.
Brain++ enables business enterprises to overcome these issues, said Sun Jian, chief scientist at Megvii. The deep learning platform not only facilitates rapid algorithm development but also makes large-scale algorithm training possible. Sun added that the deep learning platform can build custom algorithms to suit the needs of companies from diverse sectors.
According to the company, Brain++ can shorten algorithm development time by 80% and reduce the overall cost of algorithm production by 55%.
Although Brain++ was formally launched in the market in 2020, its product development began much earlier. “We started developing Brain++ in 2014,” said Tang Wenbin, co-founder and CTO of Megvii. “The goal was to allow R&D personnel to have access to a wide range of technical capabilities, from data to algorithm industrialization, so they do not have to reinvent the wheel and can easily accelerate the deployment of AI.”
By incorporating automated machine learning in its system, Brain++ can “employ algorithms to train other algorithms, and uses AI to create AI,” said Tang.
For any deep learning framework to operate optimally, it requires large amounts of unstructured data that must be manually analyzed and labeled. This process of making the data content available in various formats such as text, videos, and images that are recognizable to machines, also known as data annotation, can be time-consuming and expensive as many AI systems today require supervised learning. Faulty data can translate into biases and result in poor predictions by AI.
According to Sun, Brain++ significantly improves the speed of data annotation for deep learning, reducing the high overhead costs associated with the process of collecting and labeling data. Data annotation constitutes a major cost component in AI project development and is widely considered a top challenge faced by AI companies in the process of commercializing AI technologies.
While the rise of AI technologies can generate new growth opportunities for AI providers to create additional value for their customers, it is crucial for these firms to understand the unique set of issues that comes with commercializing AI, especially the need to develop commercially viable AI business models.
When Megvii first developed Brain++, the platform was used for internal R&D. But the AI firm’s top management soon recognized the need for the company to focus on the commercial application of its deep learning technologies to drive future growth and build organizational resilience.
This was a major reason for Megvii’s ventures beyond its core business into China’s logistics sector, tapping its AI-powered systems and robotics to fully optimize the country’s massive delivery ecosystem.
On the back of its ambition to become “the hardest AI company” operating in logistics, in 2020, Megvii launched Hetu 2.0, an update of its robotic operating system Hetu, which at the time had been used in more than 100 warehouses by different companies since its launch in January 2019.
To build wider support for its proprietary technology and boost the company’s competitiveness, in the same year, the startup decided to open-source its deep learning framework MegEngine, one of the three components of Megvii’s hallmark platform Brain++. This move aimed to accelerate the development of a thriving ecosystem around Brain++, one that would challenge the world’s two top deep learning frameworks, Google’s TensorFlow and Facebook’s Pytorch.
MegEngine is one of the three key components of the Brain++ deep learning architecture, which was developed to train computer vision algorithms at scale. The other two proprietary components of Brain++ are MegData (a data management
system) and MegCompute (a computing power dispatching system).
Megvii’s strategy of commercializing its AI technologies is in line with China’s national plan to develop home-grown tech talent, including the need to reduce Chinese companies’ reliance on US open-source frameworks. A dependence on US deep learning frameworks by Chinese startups is widely seen as a significant gap in China’s AI ecosystem.
For example, the usage of Chinese open-source AI frameworks has yet to catch up with their US-led counterparts. According to a comparison of GitHub repositories, Baidu’s open-source deep learning platform, PaddlePaddle trails TensorFlow by a factor of eight; notably, its use is declining.
This article was adapted based on portions of a feature originally written by Ji Yusheng and published on Zhidx.com (WeChat ID: zhidxcom). KrASIA is authorized to translate, adapt, and publish its contents.