Baidu Ma Yanjun: To achieve AI technology self-reliance and self-improvement, domestic deep learning frameworks face three major problems

作为推动AI应用大规模落地的关键力量，深度学习框架的重要性日益凸显。它不仅关系国计民生的行业和领域广泛的应用，同样也对信息系统的科技安全有着决定性的意义。

On March 31st, Baidu AI Open Day "AI, I'm going! ” held the fifth offline event. At the event, Dr. Ma Yanjun, general manager of Baidu's AI technology ecosystem, systematically shared the competitive landscape in the field of deep learning, the development breakthroughs and future trends of China's self-developed deep learning framework - "Deep learning framework is in the artificial intelligence technology system. At the waist, it is connected to the chip at the bottom and the application at the top."

(Dr. Ma Yanjun, general manager of AI technology ecology, shared on the spot)

Similar to the operating system Windows in the PC era, and IOS and Android in the mobile Internet era, the deep learning framework is the operating system in the intelligent era. Together with the chip, it constitutes the infrastructure of artificial intelligence. The importance of the deep learning framework is no less than that of the chip. . In the "14th Five-Year Plan", the "deep learning framework" was included in the field of "new generation artificial intelligence" and became a cutting-edge innovative technology supported by the state.

In terms of the core AI technology of deep learning framework, Chinese enterprises must take the initiative even if they face difficulties such as high threshold and difficult ecological construction. As of December 2021, Baidu's "Flying Paddle" deep learning platform has broken through the monopoly of Google and Facebook in the Chinese market in the past, and has become the largest comprehensive market share of deep learning platforms in China. At present, artificial intelligence has entered the stage of large-scale implementation, and more and more developers and enterprises are carrying out intelligent transformation applications based on domestic deep learning platforms.

How can China's industrial intelligent transformation achieve technological breakthroughs?

Domestic deep learning framework faces three difficulties

Deep learning frameworks are making AI applications easier. Based on the deep learning framework, enterprises can develop AI applications faster and more conveniently according to the characteristics and scene needs of their own industries, and no longer need to build a foundation from 0 to 1, which greatly improves the efficiency and level of industrial intelligence.

No matter from the development of AI technology or industrial application, the deep learning framework is in a very core position. Since 2013, various R&D entities in the global artificial intelligence academia and industry have successively open-sourced their self-developed deep learning frameworks, and built an artificial intelligence open platform with the framework as the core to promote the establishment of an artificial intelligence industry ecosystem. The deep learning frameworks represented by Google's TensorFlow and Facebook's PyTorch started early and developed rapidly, occupying a dominant position in the industry.

As early as 2017, the National Development and Reform Commission officially approved the establishment of the National Engineering Laboratory for Deep Learning Technology and Application, and China's deep learning framework gradually broke through the international competition. In 2021, the IDC report shows that Baidu "Fei Pao", China's first open source and open deep learning platform, has surpassed other international giants in the comprehensive share of China's deep learning market, becoming the first in China. This makes my country's artificial intelligence technology developers and users do not have to rely on foreign platforms, and can further rely on domestic platforms to cultivate industrial ecology.

However, there is still a long way to go for China's self-developed deep learning framework to take the lead in international competition. Ma Yanjun pointed out that the current development of China's deep learning framework still needs to break through three key points: technical strength, functional experience, and ecological scale.

First of all, in terms of technological innovation, the research and development of deep learning frameworks requires underlying technical talents in the field of artificial intelligence, and my country's reserves in this field are still insufficient.

Secondly, in terms of application experience, since China is the country with the most complete global industrial chain and complex industrial system, the transformation needs of small and medium-sized enterprises are imminent. However, in the process of applying AI and promoting the intelligent transformation of enterprises, it takes at least 3-6 months from the laboratory to the industrial implementation of only one technology application. A development platform with low or even zero threshold is extremely important.

In terms of development and application ecology, deep learning is a typical co-creation technology field. Only by building its own ecology can continuous iteration and development be achieved. However, the construction ecological cycle is long and the cost is high, and only when the technology and functional experience of the domestic framework are sufficient to meet the needs of developers, can there be an opportunity to cultivate an independent innovation AI development and application ecosystem.

The deep learning framework may determine the AI industry pattern in the next 5 years

Baidu Flying Paddle has become the No. 1 in the Chinese market

In the field of global deep learning, foreign developers mainly develop, train and deploy artificial intelligence algorithms and models based on foreign deep learning frameworks such as TensorFlow, PyTorch, and MxNet. The deep learning frameworks developed by Chinese AI companies still have a certain gap in terms of community prosperity and the number of developers.

However, the Chinese deep learning framework represented by Flying Paddle is developing into an open source and open platform that is more suitable for industrial needs and more popular with Chinese developers. On the one hand, China's deep learning framework has been continuously rooted in practical application scenarios, firmly grasping the needs of developers and enterprises for intelligent upgrades, and lowering the application threshold of artificial intelligence technology. On the other hand, China's deep learning framework is deeply adapted and integrated with more chip manufacturers, forming a soft-hard synergy advantage.

"Chinese enterprises and industries have their own characteristics. For example, in the fields of industry, agriculture, logistics, finance, etc., Chinese enterprises' needs for AI technology also have their own unique characteristics. If the domestic deep learning framework can meet the needs of Chinese industries in a large number of functions. At the same time, it has a low threshold and is simple and easy to develop, so there will be a great opportunity to achieve corner overtaking on the industrial-level landing." Ma Yanjun said.

Taking Baidu Flying Paddle as an example, after repeated polishing of a large number of real production scenarios, traditional enterprises have been able to achieve high-performance development, large-scale training, and agile deployment of different scenarios and different software and hardware platforms in the intelligent transformation. More importantly, Feipao has completed the adaptation and optimization of 31 kinds of chips with 22 domestic and foreign hardware manufacturers including Baidu Kunlun Core, Huawei Ascend, Intel, and Nvidia, covering all mainstream chips at home and abroad. To help enterprises reduce costs and increase efficiency.

(Overview of the adaptation of flying paddles and chips)

As of December 2021, Flying Paddle has broken through the monopoly of Google and Facebook in the Chinese market in the past, and has become the No. 1 comprehensive market share of deep learning platforms in China. At present, the Flying Paddle platform has gathered 4.06 million developers, created 476,000 AI models, and served a total of 157,000 enterprises and institutions, covering industries such as industry, agriculture, medical care, urban management, transportation, and finance.

(Paddle panorama)

With the deepening of the current digital transformation of China's industries, the ecological layout of China's deep learning framework is "blooming and bearing fruit" in thousands of industries such as industry, transportation, energy, and cities. Taking the field of smart transportation as an example, incidents of train delays caused by foreign objects hanging on high-speed rail contact nets often occur. A small foreign object may affect the travel of millions of people. Previously, relying on traditional manual inspections required 10 to 20 track maintenance workers per line every day, which not only cost a lot of labor, but also made it difficult to ensure timely detection and processing. After some attempts, Chengdu National Railway finally developed a set of "track online intelligent inspection system" using flying paddles, which realized the all-weather intelligent judgment of track defects. A set of intelligent inspection system of flying paddles, so that the guardians of the city no longer have to wear stars and moons.

Ma Yanjun said that with the open source and open source of China's deep learning framework and the implementation of larger-scale industrial applications, the application scenarios of China's deep learning framework will be more abundant in the future, and the cost and threshold will be further reduced. At the same time, the deep learning framework will integrate and innovate with more cutting-edge industries such as scientific computing, quantum computing, and life sciences.

It cannot be ignored that China's deep learning framework is still facing difficulties such as complex adaptation and deployment, and difficult application development. Building an autonomous and controllable deep learning and artificial intelligence industry ecological road is long and long, but it may determine the AI technology in the next five years. pattern and industry level. Ma Yanjun said: "Although the deep learning framework is a high-investment, long-term, and ecological competition, it has received strategic support from the country and enterprises and is the key to opening the next AI era."