Home > News content

Baidu released "Baidu Brain 3.0", the industry's first "multimodal deep semantic understanding"

via:博客园     time:2018/7/4 15:32:05     readed:257


Baidu Create 2018 AI Developer Conference site, Baidu founder, chairman and CEO Li Yanhong played a telephone recording, a guest who is about to participate in the developer conference and Baidu AI customer service after several rounds of dialogue, raised this question. Then the answer was revealed. The customer service in the recording was not a person but an AI.

The AI ​​customer service that was mistaken for real people is one of Baidu's AI capabilities this year. It is behind Baidu's natural language understanding and speech recognition and synthesis technology. At the meeting, Baidu Brain 3.0 was released. “The core of Baidu Brain 3.0 is & lsquo; multimodal deep semantic understanding & rsquo;”, Baidu senior vice president, AI technology platform system chief executive Wang Haifeng said, “Baidu Brain 3.0 has opened more than 110 leading AI capabilities.

“From the day of its establishment, Baidu began to develop and apply artificial intelligence technology,” said Wang Haifeng. Eight years ago, based on years of technology accumulation, Baidu began to fully deploy AI and officially released in September 2016. & ldquo; Baidu brain & rdquo;.

Today, Baidu's brain is constantly improving, from 1.0 to 3.0. Wang Haifeng introduced that Baidu Brain 1.0 completed the basic capacity building and core technology initial opening, 2.0 formed a complete technical system, opened more than 60 AI core capabilities, the core of 3.0 is “multimodal deep semantic understanding”, while opening 110 Multiple AI capabilities.


“Multimodal deep semantic understanding” refers to deep and multi-dimensional semantic understanding of multimodal data and information such as text, sound, pictures, video, etc., including data semantics, knowledge semantics, visual semantics, and speech semantics. Semantic and natural language semantics and other aspects of semantic understanding techniques. Wang Haifeng said, “Multimodal deep semantic understanding can not only make the machine understand and understand, but also deeply understand the meaning behind it, deeply understand the real world, and better support various applications. ”

Wang Haifeng introduced that data semantic technology can form a multi-dimensional spatial big data in multi-dimensional, heterogeneous and multi-modal worlds into a huge data semantic network containing hundreds of billions of nodes and trillions of relationships, summarizing rules, refining knowledge and discovering Value, helping economic and social development.

For example, in the intelligent operation and maintenance of new energy charging piles, combined with Baidu's big data, deep learning and other technologies for equipment monitoring, fault diagnosis, etc., can significantly improve efficiency and save costs. In terms of multi-semantic knowledge, Baidu has built a huge knowledge map containing hundreds of millions of entities and hundreds of billions of facts.

In addition to the basic entity maps consisting of entities, attributes, and relationships, Baidu also constructs maps of interest points, event maps, multimedia maps, and industry knowledge maps for different application scenarios and knowledge forms. All of this knowledge forms the basis of Baidu's brain.

Visual semantics allows the machine to understand the video from the point of view and extract the structured semantic knowledge. Visual Semantic technology is applied to World Cup video analysis. It can fully identify players, referees, balls, and people, objects and scenes in the video. It can capture events such as shooting, scoring, corner kick, free kick, substitution, etc. . Based on these semantic knowledge, it can complete the automatic interpretation of the robot, as well as highlight collections and statistical analysis of various data.

In the real life supermarket shopping scene, Baidu's visual semantic technology transforms the digitized video into structured semantic knowledge by recognizing people, actions, items and associated time series, which can realize the customer's shopping in the unmanned supermarket. A complete experience can also help store operators analyze and optimize store operations.

Speech semantic integration and natural language understanding technology enable machines to accurately identify and understand what people are saying and achieve more natural human-machine dialogue. Wang Haifeng told Baidu map a long list of tongue-worn navigation needs. Baidu map voice intelligent assistant perfectly recognizes and gives the best route. Li Yanhong’s intelligent customer service call to the participants is also behind the clip. These leading AI technologies are supporting.

According to Wang Haifeng, the accuracy of Hand-free speech recognition in Baidu's high-noise environment has increased by 10%; the speech semantic integration technology has improved the accuracy of far-field speech recognition by 10%; in speech synthesis, the emotional speech synthesis of WaveNet+ stitching Technology has greatly improved fluency and naturalness.

Taking the dialogue understanding and reading comprehension as an example, Wang Haifeng introduced the leading Baidu natural language understanding technology. Baidu's dialogue understanding technology has accumulated for many years, and by developing the latest deep attention matching model, it has increased by 4.1% compared with the best known results. In reading comprehension technology, Baidu brain has read hundreds of billions of articles, equivalent to the collection of 60,000 Chinese National Libraries, and thus accumulated knowledge of billions of entities and hundreds of billions of facts.

“By continuous acquisition and accumulation of knowledge, Baidu’s understanding of the brain is constantly upgrading, and the level of intelligence is significantly improved, which in turn can better serve users. & rdquo; Wang Haifeng said.

Baidu Brain 3.0 proposes “multimodal deep semantic understanding”, PaddlePaddle is the basis behind its technological breakthrough. PaddlePaddle is a deep learning framework independently developed by Baidu and is a deep learning framework for Chinese people. Wang Haifeng officially announced PaddlePaddle 3.0, including the complete core framework, as well as AI Studio, AutoDL, EasyDL and other platforms that allow developers to access top AI capabilities equally and easily.

According to Wang Haifeng, the PaddlePaddle3.0 core framework is fully optimized for server and mobile versions, and can be applied to a wider range of development needs. The release of the three platforms allows developers to acquire top AI capabilities more equally and conveniently.

Among them, AutoDL can search the neural network structure more efficiently and automatically, developers can get high-quality models quickly without special hardware equipment; EasyDL can help developers to customize the model of zero-based algorithm training business, operate visualization, no need to understand deep learning; AI studio has cloud Integrated, easy to use, efficient operation and free resources, it is a PaddlePaddle training platform that integrates “data, algorithms, computing power”, and meets the needs of user learning, technology advancement and academic research.

In addition, Baidu Brain 3.0 first incorporated the chip into the technical system, which enabled Baidu's brain to have a more complete integration of software and hardware, driving the explosive growth of Baidu's brain. Baidu's self-developed China's first cloud full-featured AI chip "Kunlun" also made its debut at the conference. “Kunlun” is specially optimized for speech, natural language processing, images, etc., with a 10x reduction in cost under the same performance, and high ease of use. “The AI ​​chip will be deeply integrated with Baidu's self-developed PaddlePaddle deep learning framework to promote the rapid development of the AI ​​industry. & rdquo; Wang Haifeng said.

As Baidu's brain continues to open up, more and more industries and businesses are becoming more intelligent. Today, Baidu brain calls more than 400 billion times a day. The callers include both AI engineers and zero AI” basic beginners, as well as from all walks of life who want to use AI to innovate their business and upgrade their business. Business.

“It is better to teach people to fish than to teach people. We develop the best AI technology and we are committed to opening up the best AI technology. & rdquo; Wang Haifeng said. Up to now, Baidu has opened more than 110 leading AI scenes capabilities and solutions, and has reduced the threshold of AI applications by opening up customized platforms such as EasyDL and AI capabilities of software and hardware, helping developers and enterprises to implement AI services. Innovation and upgrades.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments