On the afternoon of October 17, the 2019 ZGC Forum- Deep Learning Technology and Application Innovation Forum in AI Era sponsored by Baidu was held in Beijing. As the only parallel forum focusing on deep learning in this ZGC Forum, experts and scholars from well-known institutions and enterprises at home and abroad such as Tsinghua University, University of Maryland, Intel and Lenovo gathered in Beijing to discuss the technological frontier of deep learning and future industrial development trends.

"Deep learning development and industrial application is a complex system. Baidu has opened  a self-developed deep learning platform-PaddlePaddle. We look forward to working with all circles in the field of software and hardware to better standardize, automate and modularize deep learning technologies, promote industrial intelligence, and contribute China's strength to a better intelligent era." Wang Haifeng, chief technical officer of Baidu and director of the State Engineering Laboratory for Deep Learning Technology and Applications, said in his speech.

Wang Haifeng, chief technical officer of Baidu and director of the State Engineering Laboratory for Deep Learning Technology and Applications

Mature technology, speeding up the application and China's deep learning need to cultivate strong roots.

Deep learning is one of the fastest growing fields of artificial intelligence in recent years. Driven by three carriages of computing power, data and algorithm, deep learning has made substantial progress on the issue of speech recognition, machine vision, natural language processing and so on, thus AI entered the real application scene and began to play its real value. In the 2019 government work report, special emphasis is placed on "expanding' intelligence plus' to empower the transformation and upgrading of the manufacturing industry", which is also the third consecutive year that artificial intelligence has appeared in the government work report.

Xu Xinchao, member of Party group and deputy director of the Beijing Municipal Science and Technology Commission

Deep learning towards large-scale industrial application has become a consistent direction from policy orientation to industry consensus, so it is an important topic to build a solid base at present for this. Xu Xinchao, member of Party group and deputy director of the Beijing Municipal Science and Technology Commission, pointed out in his speech that faced with a new round of technological revolution and industrial transformation, Beijing has organized leading enterprises such as Baidu and other academic circles to take the lead in building an open AI ecosystem, focusing on basic theoretical research and research and development of key common technologies, supporting the development of standardization of open source algorithm frameworks, conducting artificial intelligence benchmark tests and soft and hard adaptation studies, and promoting the opening of application scenarios and data.

Jiang Guangzhi, deputy inspector of Beijing Municipal Bureau of Economy and Information Technology, said that the deep learning framework combined with computing chips would form the core technology system of leading industrial ecology. We are very glad to see that the domestic open source and deep learning technology system, represented by the PaddlePaddle, has initially been able to support the development of the artificial intelligence industry. In the next step, Beijing will focus on building an artificial intelligence technology system around the self-developed deep learning framework, and relevant work is in continuous progress.

Dai Qionghai, academician of Chinese Academy of Engineering and professor of automation department of Tsinghua University, attended the meeting and gave a keynote speech. China's development in the field of deep learning shall not only draw nutrients and make positive contributions to the world's open source spirit, but also focus on growing a strong roots system under its own soil. Dinesh Manocha, professor, University of Maryland-College Park also said knowledge from many fields and a large number of experiments and training have been applied in the research of robot technology and automatic driving technology, among which an efficient and open framework is very necessary.

Dinesh Manocha, professor, University of Maryland-College Park

In this forum, the fields of computing power represented by deep learning common technology platforms and chips are the two most talked about directions by scholars and enterprises at home and abroad, which are also the two most important basic links in the process of deep learning of large-scale industrialization.

Lower barriers Higher efficiency Open source platform + intelligent chip detonates industrial intelligent application

How many steps does it take from the code to the familiar face recognition, intelligent conversation and personality recommendation? The chain is very long, but its beginning is undoubtedly on the deep learning platform. Baidu CTO Dr. Wang Haifeng once compared the deep learning platform to "a foundation for all artificial intelligence applications". From networking, training, to prediction, the deep learning platform seals and packages the underlying language and important algorithm models, greatly reducing the research and development threshold, and is a typical common technology platform.

On the forum site, Wu Tian, executive director of Baidu AI technology platform system and deputy director of National Engineering Laboratory for deep learning technology and application, introduced Baidu deep learning platform PaddlePaddle and its industrial practice. In the global open source framework array, PaddlePaddle is the first and only industry-level deep learning platform with open source and complete functions in China. So far, the PaddlePaddle deep learning platform has served more than 1.5 million developers, with more than 65,000 enterprise users and 169,000 models released on the customized training platform. In addition, Wu Tian also released the White Paper on AI Technology Achievements of Baidu Brain, which fully showed the technological evolution of Baidu Brain in the past year to the industry. 

Wu Tian, executive director of Baidu AI Technology Platform System and deputy director of National Engineering Laboratory for deep learning technology and application.

In the direction of computing power, with the continuous advancement of industrial intelligence, the network structure of deep learning models, especially industrial-level models, is becoming more and more complex, and the demand for large-scale deep learning computing is surging. Assaf Schuster, director of the Artificial Intelligence Center of Israel Institute of Technology, shared the leading basic research on AI high-performance computing from the algorithm level. In the field of hardware, customized high-performance chips and machines specially designed for AI workloads and built in cooperation with software have always been widely discussed in major conferences.

Karthikeyan Vaidyanathan, deep learning multi-chip performance architect of Intel AI Products Group introduced Intel Nervana Neural Network Training Processor (NNP-T). The processor, developed in collaboration with Baidu, accelerates large-scale distributed training, which is up to 10 times stronger than similar products. Liu Daofu, vice president of Cambricon, a well-known pioneer in the field of intelligent chips in China, shared the unique thinking of Cambrian on intelligent processor design. In addition to the end-side chips, Zhang Qing, AI chief architect of Inspur Group, shared the deep learning computing optimization and application practice from the server perspective. He said that to improve the performance and efficiency of the computing system, it is necessary to consider comprehensively from the system perspective, Co-Design of training and reasoning platforms and algorithms, and application scenarios.

At the data level, Shi Zhongchao, director of artificial intelligence laboratory of Lenovo Research, stressed the importance of industry knowledge in the process of industry landing. Shi Zhongchao said that Lenovo artificial intelligence focuses on the three major directions of intelligent Internet of Things, transformation of traditional IT to intelligent infrastructure, and industrial intelligence. In the future, artificial intelligence must obtain requirements from practical applications, combining data, algorithms, and industrial knowhow (knowledge, experience, and process) to create intelligent vertical industrial solutions.

In addition to excellent thematic reports, Baidu also invited Cheng Jian, Researcher of State Key Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, doctoral supervisor, director of the Joint Laboratory of Artificial Intelligence and Advanced Computing, Shi Zhongchao, director of the Artificial Intelligence Laboratory of Lenovo Research, Yang Ruigang, director of Baidu Robot and Autopilot Laboratory, Liu Xianglong, associate professor and doctoral supervisor of the School of Computer Science and Engineering, Beihang University, and Karthikeyan Vaidyanathan, deep learning multi-chip performance architect of Intel AI Products Group, five leading scholars and industry leaders from home and abroad, conducted in-depth discussions on "the technology trend of deep learning and application landing".

Artificial intelligence is changing from being driven by academia to being driven jointly by academia and industry, which requires collaborative innovation of IUR (Industry-University-Research). Baidu also looks forward to working with all sectors to promote technological progress, share technological achievements, promote industrial intelligence and contribute China's strength to global economic development and social progress.