Deep Learning Framework is an operating system in the Intelligent Age,Explain the Intelligence of Baidu Flying Propeller Industry in detail

On June 28, the China Artificial Intelligence Summit (CAIS 2019) was held in Nanjing. With the theme of "Intelligence + Embrace the Future", the conference brought together many leaders and representatives in the field of artificial intelligence to discuss the present and future of China's artificial intelligence. Wu Tian, Executive Director of Baidu AI Technology Platform System and Deputy Director of National Engineering Laboratory of Deep Learning Technology and Application, was invited to attend the "AI Open Source and Open" Summit Forum and delivered a keynote speech. The paper expounded the current situation of open source software in the world, as well as the core advantages of Paddle Paddle, the first self-developed Deep Learning Platform in China, and the practice and results of promoting industrial intelligence.

Open source is one of the hotspots of AI. It promotes the prosperity of global AI technology and industrial ecology, and becomes an important driving force of industrial intelligence. Wu Tian introduced that the global open source software is on the rise, and China's contribution to it is growing. Open source software is accelerating globalization, and collaboration is also tending to globalization; the rapid development of artificial intelligence promotes the prosperity of open source software ecosystem; in the open source community, the contribution of related open source projects of companies and institutions is increasing; open source software has become an important mode of software and interconnection development. "Open source and open platform are important for the deep integration of AI and industry," Wu Tian said. The combination of AI technology and industry produces value, which involves key elements such as multi-technology integration, capacity reuse, industry chain connection, and industry and national security. Obviously, the self-developed open source platform of artificial intelligence is an important force to continuously promote the development of artificial intelligence, accelerate the process of industrial intelligence, and promote scientific and technological innovation and industrial change in China.

"This upsurge of artificial intelligence is inseparable from the continuous breakthrough and wide application of in-depth learning technology. Among them, the framework of in-depth learning is the core foundation, and its importance is self-evident. Wu Tian said that the in-depth learning framework is an operating system in the era of intelligence. There will be a lot of applications based on the operating system, which will be of great help to the innovation of artificial intelligence. Paddle Paddle, the only fully functional in-depth learning platform built by Baidu, is promoting the development of artificial intelligence and accelerating the process of industrial intelligence in China.

Wu Tian introduced that the paddle is an end-to-end open source deep learning platform which integrates core framework, tool components and service platforms. It includes models supporting real-world scenario applications and achieving industrial application effects, distributed training capabilities for large-scale data scenarios, and high-speed reasoning engines supporting heterogeneous hardware.

Firstly, the core framework layer of the propeller opens up a whole set of capabilities of development, training and prediction, and opens up more than 70 official models verified by real business scenarios, covering several key technical fields such as vision, NLP, recommendation and voice. The propeller has become the most in-depth learning platform supported by the official model. Paddle NLP (Paddle NLP), a Chinese NLP toolkit for industrial applications, implements a set of shared skeleton codes for various models in the field of natural language processing at the task level, making the network more flexible. ErNIE, which is the leading Chinese semantic representation model in Chinese NLP tasks, has been introduced in the model library, which greatly enhances the semantic representation ability of the model and has achieved remarkable results in practical application. The video recognition toolkit provided by the propeller covers the mainstream practical sequence modeling algorithm and end-to-end video recognition model, and provides a series of application technology solutions such as video classification and video location for developers.

Aiming at the large-scale sparse feature, a large-scale sparse parameter server is designed and opened for the application scenario of large-scale data, massive feature and self-expansion, high-frequency model iteration. At the same time, for industrial data processing, the propeller opens up distributed sample Shuffle, distributed file system IO support, flexible embedding of multi-language IO components and other capabilities.

Model deployment is an important step for deep learning and application. Flying oars provide a complete end-to-end full process deployment solution, and continue to expand support for various hardware. On this basis, the paddle provides a leading performance of the underlying acceleration library and reasoning engine. The recently released Paddle Serving library supports rapid deployment of the server side.

At the tool component level, the propeller provides tools and components such as Paddle Hub, PARL and AutoDL Design to meet the needs of developers to develop in-depth learning applications flexibly and efficiently. Among them, the design component of the network structure of the propeller automation is based on the deep learning design, which has surpassed the effect of human expert design in an all-round way.

In the service platform, the paddle provides zero threshold customized training and service platform EasyDL for developers without deep learning technology foundation. It helps developers realize automatic processing and training based on customized data such as pictures, dialogues and sounds, and quickly complete deployment. In addition, one-stop training development platform AI Studio provides developers with a more perfect and convenient programming environment; EasyEdge end computing model production platform supports a variety of in-depth learning frameworks and network structures, and can adapt to most mainstream chips and operating systems.

Flying oar, an open source and deep learning platform originating from industrial practice, has been playing a more and more important role in the process of industrial intellectualization in all walks of life. Taking "AI Recruiter" as an example, Wu Tian introduced the cost-benefit of the personnel-post intelligent matching system based on flying oars for enterprise recruitment. "AI Recruiter" makes full use of the model advantage of the flying oar in the field of Chinese NLP, and can make better matching between personnel and posts. The HR invitation interview of the enterprise is successful.

Please read the Chinese version for details.