The hottest Wang Haifeng attends the new generatio

  • Detail

Wang Haifeng attended the new generation of AI academician Summit Forum: Baidu AI's open source and opening up

December 18 - on December 17, the new generation of AI academician Summit Forum hosted by Pengcheng laboratory was held in Shenzhen. At the meeting, Wang Haifeng delivered a keynote speech on "Baidu AI open source and opening", explaining the development path of the new Ingeo formula to open source and open up the "high temperature resistant" conventional artificial intelligence technology

Wang Haifeng believes that open source is crucial for the era of artificial intelligence. On the one hand, the development and breakthrough of artificial intelligence technology requires everyone to work together and promote it, so open source and openness are particularly important; On the other hand, AI technology will empower all walks of life. Open source and openness will make it easy for all walks of life to use AI, so that all sectors of society will benefit even if the rest of the society can't continue to transfer the load

artificial intelligence is highly valued by the country and is regarded as an important driving force for the new round of scientific and technological revolution and industrial change. Promoting the deep integration of artificial intelligence and social and economic development is the focus of the current development of artificial intelligence

Wang Haifeng said that AI technology has been widely used in the Internet field. For example, search engines that directly present answers to users, intelligent speakers that interact with users through voice, maps that provide path planning for users to travel, and so on. Behind these interconnected applications, natural language processing, knowledge maps, voice, image, deep learning and other industrial intelligence technologies are playing a role

Baidu has accumulated in the field of artificial intelligence for nearly 20 years. Since 2000, Baidu's search business has its own artificial intelligence gene. For example, when understanding user needs and page content, it needs to be realized through natural language processing technology; In 2010, Baidu has foreseen that AI will play a qualitative role in promoting the development of the Pentium industry for the future performance of electronic omnipotence, and began to make a comprehensive layout; In 2016, baidu announced the opening of AI core capabilities

after years of accumulation and precipitation, Baidu has formed a complete AI technology layout including the basic layer, perception layer, cognitive layer and platform layer, including basic algorithms, data and computing power, perception technologies such as voice and image, cognitive technologies such as natural language processing, knowledge atlas and user understanding, as well as an AI technology open platform that is open to the outside world and opens AI core technologies

deep learning framework is the underlying foundation for the current AI technology to play a role. It is equivalent to the "operating system" of AI architecture. According to Wang Haifeng, Baidu has started the research on deep learning since January 2012, and has successively applied deep learning to search and other businesses, achieving good results. In August 2016, Baidu paddepaddle was officially open source, becoming the first and only open source, complete and complete in-depth learning framework in China, and has been widely used and verified in Baidu feed flow recommendations and other scenarios

what does a complete and comprehensive in-depth learning framework include? Wang Haifeng interpreted Baidu's deep learning full-function suite paddepaddle suite, which is composed of a technologically leading core framework, rich supporting modules and components suitable for enterprise applications, and a service platform that reduces the threshold of application

"the key to this complete and comprehensive in-depth learning framework is still the core framework," Wang Haifeng said. The core framework layer includes paddepaddle fluid training and prediction framework. At the same time, there are such system prediction engines for server and mobile terminals, and open source special models for different AI applications, such as intelligent text processing, intelligent vision, intelligent recommendation, etc. Paddlepaddle core framework has five characteristics, which can support large-scale heterogeneous computing clusters and large-scale parallel training - paddlepaddle supports parallel training of dense parameters and sparse parameters at the same time, with sparse parameters up to 100 billion, and provides a variety of algorithms, uses multi-channel communication, and is suitable for multi terminal deployment

flexibility, simplicity and ease of use are developers' demands for deep learning framework. Based on this, Baidu has developed a series of auxiliary tools and service platforms such as autodl, visualdl, Parl, easydl, AI studio, etc. Autodl supports deep learning design, migration and adaptation. The model generated by autodl automatic design is even better than some typical models designed manually; Visualdl can realize the deep learning, training and prediction of the whole process visualization, and help developers better control the overall situation; Parl reinforcement learning tool helped Baidu win the championship with great advantages in nips 2018 AI prosthetic challenge; AI studio, a one-stop development platform, can provide developers with a more perfect and convenient programming environment; Easydl helps users realize automatic processing and learning based on customized image data, dialog data, sound data, etc., quickly complete deployment, and reduce the threshold of use. Taking the automatic detection of floor defects as an example, Wang Haifeng introduced the convenience brought to developers by the paddepaddle related platform. "No one in this company who produces floors understands deep learning at all, but based on easydl platform, they upload defective floor data, and they can automatically generate data models to automatically detect floor defects, which has more than doubled the daily processing volume of floor quality inspection."

the open source of the underlying deep learning framework is only a small part of Baidu's layout of AI and development of AI technology ecosystem. Wang Haifeng said that in order to let AI penetrate all walks of life and enter thousands of households, Baidu has fully opened more than 140 AI core technologies, such as voice, vision, natural language processing, knowledge atlas and so on. For example, in the field of natural language processing, Baidu has not only opened up a large number of classic models of natural language processing based on paddepadder reconstruction, but also fully opened up Baidu's independently developed lexical analysis, semantic matching, emotion analysis, topic model, dialogue system, etc. In terms of vision, voice and other technologies, baidu continues to open source and facilitate the reform and innovation of industry applications

let every developer have equal and convenient access to AI capabilities is the driving force and vision of Baidu AI open source. Nowadays, more and more developers and partners benefit from the open source and open Baidu AI technology, accelerating the transformation and upgrading of the industry

Copyright © 2011 JIN SHI