China has grand ambitions to become a global leader in artificial intelligence. In its 2017 “Next Generation AI Development Plan,” the country outlined its strategy to be the top AI power in the world by 2030. As China races for technological supremacy, new homegrown innovations have been observed, especially in deep learning, catalyzed by the advent of big data and advancements in computing power.
One such innovation is Brain++, a proprietary AI productivity platform, created by Megvii, one of China’s four “AI dragons.” The deep learning framework helps businesses build their own AI capabilities, deploy customized algorithms at scale, and accelerate their digital transformation. Since the development of Brain++ began in 2014, the AI framework has gone through eight iterations.
Established in 2011, Beijing-based Megvii is a developer of deep learning technologies and is known as the creator of facial recognition software Face++, the world’s largest open-source computer vision platform. Much of the technological breakthroughs achieved at Megvii in recent years can be credited to the late Sun Jian, a former Microsoft researcher who specialized in computer vision and computational photography.
Sun joined Megvii in 2016 as managing director of research. As the company’s chief scientist, Sun built algorithms that laid a foundation for Megvii’s advanced deep neural networks that can be deployed across cloud, mobile, and edge computing platforms to achieve better performance in algorithm accuracy and training speed.
The AI scientist was known for his groundbreaking research on ResNet, a cloud-based deep neural network that he co-developed earlier and was widely considered a major breakthrough in image processing technology.
At Megvii, Sun was also instrumental in the creation of ShuffleNet, a mobile-based deep neural network designed for devices with relatively limited computing power such as mobile devices, as well as DorefaNet, a lightweight deep neural network model suitable for chips.
The evolution of deep learning frameworks
Deep learning has come a long way since the idea was first conceived in the 1950s. Over the past decade, the AI industry has undergone rapid evolution, leading to new learning paradigms.
2012 was a breakthrough year for deep learning with the development of AlexNet, a complex multi-layer neural network that can recognize visual patterns directly from pixel images with minimal preprocessing. It was around this time when early deep learning models such as Caffe emerged.
The launch of Google’s TensorFlow in 2015 marked another milestone in AI framework development. The Python-friendly open-source library for numerical computation allows business enterprises to feed data into AI models and train their systems without having to create their own from scratch.
Characteristic of a next-generation deep learning framework, whereby algorithms are represented as computational graphs, TensorFlow uses graph structures to express and evaluate mathematical expressions. Computational graphs divide complex computations into small, executable steps that can be performed quickly.
According to Sun, Brain++, which also leverages computational graphs to derive algorithms for its deep neural networks, was developed even earlier than TensorFlow.
When Megvii first created Brain++ in 2014, it was initially developed for internal R&D purposes. To generate future business growth, in 2020, the tech firm brought the AI framework to the market for developers building AI solutions for commercial and industrial use.
Unlike the AI frameworks developed in North America that were initially created for research, such as Caffe and Theano, many of China’s deep learning platforms such as Brain++ were designed to solve business problems.
This explains Megvii’s business-oriented approach in the development of its AI technologies. For example, the AI company significantly improved the speed of data annotation, which benefits AI developers by reducing their high overhead costs associated with the process of collecting and labeling data.
Deep learning-based systems are highly dependent on data for their predictive power. The open science nature of AI means that firms’ competitive advantages often stem from the extent to which they can assemble a large database faster than anyone else.
Aware that big data is crucial to AI innovation and for optimizing training performance, Megvii endeavored to invest in new data sets and improve data quality. Early this year, in collaboration with nonprofit Beijing Academy of Artificial Intelligence, the AI firm built Objects365, one of the world’s largest, fully annotated object detection datasets in terms of its number of images. The dataset contains 365 object categories, over 638,000 images, and more than 10 million annotated bounding boxes.
At the same time, Megvii incorporated multitask learning in its deep learning architecture, allowing a set of tasks to be solved at the same time to improve learning efficiency. During a Series C financing interview in 2017, Megvii co-founder Yin Qi said, “We are making every effort to develop an all-in-one system.”
By designing Brain++ as an end-to-end deep learning model, as opposed to a components-based framework, hundreds of researchers can conduct tens of thousands of training tasks on thousands of graphics processing units simultaneously through the platform.
Challenges for Megvii
With current operations in three main business segments—consumer IoT, city IoT, and supply chain IoT—Megvii appears to have achieved some measure of success in leveraging its proprietary AI technologies to drive strategic growth.
But much of Megvii’s growth can be attributed to the Chinese government’s AI-driven policies in building smart cities across China. Heavy reliance on domestic government contracts raises questions about the company’s ability to become a truly global AI player.
Additionally, despite seeing revenue growth in recent years, the company has failed to turn a profit, largely due to significant investments in R&D. The AI firm posted a loss of RMB 6.64 billion (USD 1 billion) in 2019 and RMB 2.85 billion (USD 437.9 million) in the first three quarters of 2020. In the first half of 2021, Megvii also reported a loss of RMB 1.86 billion (USD 268 million), according to the company.
Megvii faces even more challenges ahead. The Chinese AI firm has to contend with tough competition from global players such as Meta’s PyTorch and Google’s TensorFlow, deep learning frameworks that have been widely adopted around the world.
On the domestic front, Brain++ competes with Chinese tech goliaths, including Huawei’s MindSpore and Alibaba’s X-DeepLearning, both of which have access to more resources such as funding, tech infrastructure, and talent.
A major domestic rival is Baidu’s PaddlePaddle. Launched in 2016, PaddlePaddle is China’s first open-source deep learning platform in China. The framework is currently used by 4 million developers and over 157,000 enterprises from a myriad of industries.
According to global market consultancy IDC, PaddlePaddle ranked first in terms of market share among deep learning platforms in China as of December 2021, surpassing TensorFlow and PyTorch.
This article was adapted based on portions of a feature originally written by Ji Yusheng and published on Zhidx.com (WeChat ID: zhidxcom). 36Kr Global is authorized to translate, adapt, and publish its contents.