Shulou(Shulou.com)06/02 Report--

2019-10-19 20:52:04

Technological change has never been dramatic, when you suddenly realize that the world has been turned upside down by deep learning, the fastest-growing technology of artificial intelligence in recent years.

The 2019 Zhongguancun Forum on Deep Learning Technology and Applied Innovation in the AI era, hosted by Baidu, was held in Beijing.

Experts and scholars from well-known institutions and enterprises at home and abroad, such as Baidu, Tsinghua University, University of Maryland, Intel, Lenovo, etc., gathered together to discuss the cutting edge of deep learning technology and the future trend of industrial development.

During this forum, Baidu released for the first time "Baidu brain AI Technology achievements White Paper", showing the technological evolution of Baidu brain over the past year. Wu Tian, executive director of Baidu AI technology platform system and deputy director of the National Engineering Laboratory for Deep Learning Technology and Application, also declassified the four leading technologies of Flying Propeller on the spot, fully demonstrating a series of achievements of this standardized, automated and modular deep learning platform.

Development History of ▲ artificial Intelligence Technology

In the past 60 years, the development of artificial intelligence has experienced three typical stages: artificial rules, machine learning and deep learning. Around 2012, deep learning algorithms have made great breakthroughs in the field of computer vision and speech recognition. Driven by the troika of computing power, data and algorithm, deep learning has set off the third upsurge of artificial intelligence. This new change involves the Internet of things and reshapes the core of tens of thousands of hardware devices, which is not only the focus pursued by technology companies, but also increasingly become the inevitable choice for innovation and change in all kinds of traditional industries.

From Baidu Flying Propeller, the "handle" of the domestic open source deep learning framework, we can see the epitome of the development of domestic deep learning, the evolution of deep learning technology and applications, and the changes it has brought to China's artificial intelligence industry. at the same time, it also discusses the crises and opportunities faced by China in the R & D and landing of deep learning technology.

Occupy the policy highland for one or three years and usher in a new era of domestic AI

This year is the third year in a row that artificial intelligence has appeared in the government work report, and it is also the first time that the government work report has put forward the concept of "intelligence +".

According to the planning of the State Council, the scale of China's artificial intelligence core industry will reach 150 billion yuan in 2020, and will continue to maintain rapid development in the next decade. At present, this value in China has exceeded 100 billion RMB, and success is imminent to achieve this goal. According to the data disclosed by the vice premier at this year's Zhi Expo, according to preliminary statistics, the scale of industries related to artificial intelligence may have exceeded 500 billion yuan in 2018 and is becoming an important new economic growth point.

Combined with the development of artificial intelligence in China in recent years, the national planning of artificial intelligence focuses on technological research and development and industrial development.

The research and development of artificial intelligence in China started early, and there has been a basis for related research in academic circles since 1977.

In industry, Baidu has been building large-scale machine learning infrastructure, models, tools and experimental platforms since 2008. In 2012, Baidu successfully applied the deep learning DNN model to speech recognition and OCR tasks, and in 2013 created the prototype Paddle of the deep learning framework PaddlePaddle. In 2016, Baidu officially opened up PaddlePaddle, making it the first and only open source and fully functional deep learning framework in China.

Wang Haifeng, chief technology officer of Baidu and director of the National Engineering Laboratory for Deep Learning Technology and applications, once compared the deep learning framework to "the operating system of the intelligent era". The deep learning framework plays the role of connecting chips and inheriting various applications.

Wang Haifeng, Chief Technical Officer of ▲ Baidu and Director of the National Engineering Laboratory for Deep Learning and applications

Nowadays, artificial intelligence in China has gone through the embryonic stage and gradually entered the rapid landing period. As the infrastructure of the artificial intelligence era, the deep learning framework is gradually integrated and infiltrated with the industry, which has a far-reaching impact on people's production and way of life.

Second, open source deep learning framework to catalyze the outbreak of artificial intelligence applications.

Deep learning framework plays an important role in the history of artificial intelligence. Jiang Guangzhi, deputy inspector of economics and information technology in Beijing, said that the deep learning framework combined with computing chips will form the core technology system of leading industrial ecology.

However, at present, the mainstream frameworks such as TensorFlow, Caffe, PyTorch and MXNet in the world are in the hands of foreign technology giants such as Google, Facebook, Amazon and so on. At present, only Baidu Flying Propeller breaks through the tight encirclement and strives for a place in the ranking of global open source deep learning framework.

As of Oct. 18, the number of star on Baidu's Github has reached 10055, the number of forks has reached 2686, and the number of code updates has reached 25480.

Although foreign frameworks are easy to use, there are no Chinese-related documents and materials for a long time, and hidden dangers such as data security always exist. In addition, in terms of language understanding, it is difficult for foreign enterprises to do enough in-depth research on Chinese NLP.

The emergence of Baidu flying oars not only provides a new choice for domestic developers, but also makes Baidu itself have autonomy in artificial intelligence. Baidu Feipao's newly upgraded continuous learning semantic understanding framework ERINE2.0 in July has surpassed Google BERT and XLNet on a total of 16 Chinese and English tasks, and has learned more than 1.3 billion knowledge so far.

According to the "Market share Survey of China's Deep Learning platform" released by IDC, a well-known international data analysis company, Baidu Deep Learning platform ranks third in the overall market share of China's deep learning platform, second only to Google and Facebook, and leads the domestic platform.

Its Flying Propeller-based service platform EasyDL has high market recognition, with a user recognition rate of 46.4% and a high-frequency utilization rate of 32.7%.

Wu Tian, executive director of ▲ Baidu AI technology platform system, introduced Flying Propeller industry-level deep learning open source open platform.

As the cornerstone of Baidu artificial intelligence application, Baidu flying propeller has four characteristics: flexibility and efficiency in technology, support for ultra-large-scale deep learning model training, and high-performance reasoning engine for multi-terminal and multi-platform deployment.

Baidu Flying Propeller supports more than 100 mainstream models, while opening up more than 200 pre-training models. So far, it has attracted more than 65000 enterprise users and released 169000 models on its customized training platform alone.

The deep learning framework is also a key link in promoting the landing of the industry, and hundreds of degrees have done a lot of work in the combination with applications. Many models of Baidu flying oars have not only achieved good results in international competitions, but also been verified by the industry, covering mainstream application tasks such as vision, natural language processing, recommendation and so on. At the same time, considering the increasing domestic development needs, the ease of use and efficiency of model development and deployment on its platform are continuously optimized.

Based on this framework, Baidu internally integrates artificial intelligence into its own services, driving its own business value-added. For example, Baidu Maps uses flying oars to increase the accuracy of user travel time estimation from 81% to 86%. It is widely used in industry, agriculture, service industry and so on, serving more than 1.5 million developers, not only strengthening Baidu's own artificial intelligence ecosystem, but also promoting the integration and innovation of industries and artificial intelligence in various industries.

Third, reshape the deep learning of the whole world

Most people feel the deep learning technology, or Siri, Cortana, small degree that can be "flirted" anywhere, or can control all kinds of intelligent devices with the use of words, or machine translation is not afraid to travel all over the world, or faster and faster typing speed, more and more labor-saving text recognition and picture recognition, or more and more intelligent picture repair and beauty to save disabled parties and rookies, or it is more convenient to go to the government or bank. Or a blueprint for driverless cars and no congestion is becoming possible.

In addition to the little convenience brought to people in daily life, the technical support provided by the deep learning framework is bringing huge economic value-added space and opportunities to various industries through deep integration with the real economy.

Panoramic picture of ▲ artificial intelligence industry ecosystem

1. Give birth to new enterprises and employment opportunities

The landing tide of deep learning has spawned hundreds of artificial intelligence start-ups, providing new services and solutions such as face recognition, speech recognition and machine translation. A large number of data tagging companies have also sprung up, providing more low-threshold jobs.

According to iiMedia Research, China's artificial intelligence sector raised 63.4 billion yuan in 2017 and 131.1 billion yuan in 2018, a growth rate of more than 100 percent.

2. Drive upstream semiconductor innovation

The rapid development of industrial intelligence has led to a surge in the demand for large-scale deep learning computing, and the demand for computing power in different scenarios such as cloud, edge and end makes more and more semiconductor manufacturers, startups and cross-border players set foot in artificial intelligence chips, and different types of computing units such as CPU, GPU, FPGA, ASIC, and neural mimicry chips begin to exert their vitality.

Baidu has also joined the battle as a cross-border. In addition to creating a self-developed AI chip Kunlun to provide exclusive computing power for flying oars, Baidu also announced this year that flying oars will cooperate with Huawei Kirin NPU to strengthen the combination of software and hardware at the end-to-side layout.

3. Enable downstream hardware upgrade

Deep learning has also given birth to some new types of hardware equipment, including translation pens, transcribers, face recognition and verification machines, and intelligently modified categories such as intelligent speakers, intelligent televisions, intelligent light bulbs, robots and so on.

In recent years, the global intelligent speaker market has grown substantially, and the Chinese market has grown the fastest. The domestic market has formed a tripartite confrontation among Alibaba, Tmall, Xiaomi and Baidu DuerOS family.

4. Mining dividends in traditional industries

The main battlefield of Internet traffic dividend competition is online, and deep learning is expanding this battlefield to offline dividends through restructuring with traditional industry resources.

Looking ahead, the potential intelligent stock market is vast. Health care, finance, insurance, agriculture, retail, travel, home, education. Almost as long as the scene related to human activities, there is a lot of room for optimization.

In addition, in the field of cloud computing, public cloud platform has become an important channel for low-cost access to artificial intelligence. The leader of global and domestic public cloud market is also the leader of artificial intelligence market.

IV. The advantages and disadvantages of the development of deep learning in China.

If the industry wants to achieve considerable sustainable development, it needs to return to the importance of technology itself. While the application of artificial intelligence is in full swing, we might as well settle down and examine the advantages and disadvantages of deep learning research and industrial construction in our country.

Generally speaking, the development of deep learning in China has obvious advantages in data collection and labeling, science and engineering personnel training, landing application and so on. With the help of the open source framework, many institutions or developers in China have won the championship in international competitions or evaluations in the field of deep learning technology research, and verified the cost reduction and efficiency brought about by deep learning technology in combination with the industry.

For example, Baidu won many international competitions such as ICME facial key Point Detection Competition and Multi-target tracking Challenge MOT with the help of flying oars, and also broke Stanford University's four world records of DAWNBench. These winning models are also open source on flying oars.

At the same time, artificial intelligence technology is used to feed back data marking and chip design tools to effectively improve the efficiency of data processing, optimize the chip design process, and lay a better foundation for the development of artificial intelligence.

All these make our country walk in the forefront of the world in the technology layer and the application layer, but as mentioned earlier, in the basic layer field represented by the deep learning platform, our development is relatively late, except that there are few open source depth frameworks available in the world, and the research on basic theory and technology is not deeply rooted.

Although our research institutions and enterprises have reached the forefront of the world in the application of artificial intelligence technology, our country is still relatively lacking in major original basic theories and technological achievements. Emotional recognition, autopilot, small data learning and other cutting-edge technologies still need to be continuously studied and explored, and there is still a long way to go before the real landing application.

The next step is the small goal of artificial intelligence.

After talking so much about the past and present, we should also think about the next step of artificial intelligence in our country.

In this forum, Xu Xinchao, a member of the party organization and deputy director of the Beijing Municipal Science and Technology Commission, talked about several tasks to be carried out in the field of artificial intelligence in Beijing as a whole in his speech. these directions can also be used for reference for policy planning in industry and other regions.

Xu Xinchao, member and Deputy Director of the Party Organization of ▲ Beijing Municipal Science and Technology Commission

First, forward-looking layout, the basic theory leading to the change of artificial intelligence paradigm and the change of common technology, thinking about what to do and how to do it next.

Second, support the standardization of the development of open source algorithm framework and form a joint force. Baidu is a pioneer in this field in China. Next year, we will also carry out an open source framework, and we need to explore how to form a joint force to do it.

Third, support artificial intelligence benchmark testing and software and hardware adaptation research.

Fourth, promote the openness of application scenarios and data. One is for the government, and the other is for resources to increase the application of artificial intelligence in transportation, medical care, science and technology, the Winter Olympic Games, and so on, so as to provide a better application scene for everyone.

In addition, the establishment of laws and regulations related to artificial intelligence has been put on the agenda.

As early as 2017, the "New Generation artificial Intelligence Development Plan" issued by the State Council stressed the need to "build a more perfect system of laws and regulations, ethical norms and policies on artificial intelligence." At the opening ceremony of several artificial intelligence conferences this year, a number of national cadres have stressed the need to attach importance to the construction of laws and regulations.

In 2017, ▲ State Council issued an excerpt from the New Generation artificial Intelligence Development Plan.

This year, Western countries on the other side of the ocean have successively issued a number of laws related to artificial intelligence. Not long ago, Shenzhen artificial Intelligence Industry Association and a number of enterprises issued the "Convention on self-discipline of the New Generation artificial Intelligence Industry." the aim is to build a moral and ethical system of artificial intelligence. With the development of the next policy and plan, it is believed that the laws and regulations used to ensure the healthy development of artificial intelligence in our country are on the way.

Conclusion: grow strong roots in one's own soil.

Looking back over the past few years, the development of the domestic artificial intelligence industry is booming, and the domestic open source deep learning technology system represented by Baidu Flying Propeller has the initial ability to support the development of artificial intelligence industry.

Change is accelerating, but it is still a long way from becoming a real artificial intelligence power.

As Dai Qionghai, academician of the Chinese Academy of Engineering and professor of the Department of Automation at Tsinghua University, said, the development of China's deep learning field should not only absorb nutrients and make positive contributions to the world's open source spirit, but also pay attention to growing strong roots in its own soil.

Open source provides a steady stream of nourishment for artificial intelligence entrepreneurs, but we should not just focus on the immediate convenience and ignore the long-term considerations. With the prosperity and development of the market, the foundation of China's basic theory, technology and infrastructure construction is not solid.

This requires Baidu and more domestic artificial intelligence industry chain practitioners to pay more attention to the promotion of basic research and the construction of the core technology system, so as to realize the independent control of technologies such as deep learning, and lay the foundation for competing for the international voice of artificial intelligence technology and becoming an innovative power and an economic power.

Https://www.toutiao.com/i6749496058205700619/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.