Lei Feng network AI technology commentary news on July 3, the annual Baidu Create 2019 AI developer conference arrived as planned, in the morning after the display of Baidu speaker series update, release of the Hongqi chip, announced a number of collaborations in automatic driving In the afternoon, Baidu Brain Forum announced heavily in the afternoon.
Baidu CTO Dr Wang Haifeng said as the first guests that in the morning all the application of Baidu saw AI technology in various industries, such as graduated sound and automatic driving, all carried on the newly upgraded Baidu Brain 5.0. Not enough to see all the technological advances. This forum will provide a more detailed interpretation of the Baidu brain from the computing power to the platform to ecology.
After that, Baidu Fei played Paddle (PaddlePaddle), Baidu AutoDL, Baidu Robotics, Baidu Voice, Baidu Map, Baidu Intelligent Visual Interaction, Baidu Great Computer Technology and many other coaches one by one, giving an extensive audience to the audience. Technical party.
The latest development of Baidu aircraft platform and its application
As a deep learning platform Baidu strongly recommends, Flying Paddle is of course the focus of this forum. Ma Yanjun, head of Baidu Deep Learning Technology Platform, presented the three major advances in flying image search, target detection and industrial quality control. Later, he once again talked about the five core advantages of airplanes: full support for dynamic graphs and static graphs. There are over 70 official support models that support large-scale distributed training, providing a very strong end-to-end Developing the ability to provide systematic deep-learning technology services.
Dr. Ma Yanjun said the official model covers three important tasks. PaddleNLP currently supports more than 20 models. PaddleCV supports more than 50 models. Paddlehub offers online deployment capabilities and supports 40+ pre-training models.
He later introduced that Baidu's product system faces a number of challenges, such as terabytes of data and tens of billions of training data. Under these challenges, they created a large-scale parametric server solution. The distributed training API has also been updated in the recently released Paddle Fluid v1.5 release.
When it comes to tool development, developers are closely related to developer training. There are currently three training systems for Flying Paddle, Huangpu College, AI Fast Track and PaddleCamp. Finally, he mentioned 2019 Baidu Star, which includes Baidu Star Programming (Basic Algorithm + AI Match) and Baidu Star Development (based on the Flying Paddle Target Detection Mission). The registration for this event will be this month. At the end of the year, the national finals will be held in October this year.
Baidu released the industry knowledge map platform and intelligent creation platform 2.0
Wu Tian, Managing Director of Baidus AI Technology Platform Systems, took the latest developments in Baidu's language and knowledge technology platform. She said that natural language processing, knowledge planning and other technologies are cognitive level technology that makes it possible for machines to learn how to use human language and knowledge to be able to think about this foundation. Then she introduced Baidu's newly open source ERNIE model, which not only distinguishes itself in certain tasks in the academy, but also has very powerful application possibilities, in Baidu products such as intelligent question and answer, CTR estimation, text retouching, dialogue understanding etc. Excellent performance.
As the focus of this forum, Wu Tian officially released the knowledge map and intelligent creation platform 2.0. She said that the industry knowledge card platform is to solve the industry's need for knowledge building and knowledge application in the intelligent process. Basically, the intelligent creation platform is reorganization and reprocessing of information. Through basic techniques for natural language processing and knowledge mapping, a large amount of original information is deeply understood and logically analyzed to form the knowledge map and the material needed for creation. The library can give the creators the opportunity to mount, correct and polish the color in the written stage of the creation. Intelligent Authoring Platform 2.0 has three new features: real-time hotspots associated with rich materials, comprehensive creative help features, and leading generation technology.
Hot automatic machine learning
With the hard automatic machine learning, Baidu has taken the latest part of this technology.
Bai Jun, head of Big Data Lab, said there are three key technologies for autonomous artificial intelligence: custom networking design, adaptive and self-adapting the computing environment. AutoDL 1.0 shows that deep learning is comparable to human experts, and AutoDL 2.0 demonstrates Deep Learning has already outperformed human experts, and AutoDL 3.0 has been completely upgraded from design, migration and customization. In the design of AutoDL, Baidu proposed styleNAS (style migration with neural network structure search), which can greatly simplify the calculation and improve image quality.
He then mentioned some typical application scenarios of automated modeling, such as semantic segmentation, target detection, image classification, etc., such as garbage selection, and the camera is placed in the water to detect fish for ecological protection. In addition to a series of visual applications, he also mentioned the application of Baidu-automated modeling in NLP. In the Chinese emotional classification task, this technique achieved the same precision as expert manual setting, but greatly saved labor. .
Baidu speech technology interpretation
During the convening of this developer conference, we also saw many breakthroughs that Baidu made in voice. Gao Liang, senior manager of Baidu's voice technology department, introduced Baidu's three major technical breakthroughs in voice. He looked forward to Baidu's progress in speech recognition technology from 2012 to the present. With the continuous improvement of the model, the accuracy becomes higher and higher.
He focused on the multilevel SMLTA flow, the truncation attention model, which is the first time in the world to achieve local attention model in addition to the entire attention model, and also the first international attention model for large-scale voice usage, whether input it is still in the speaker, and the relative accuracy of the effective products can be greatly improved.
Later, he emphasized the MEITRON style migration technique, hoping that the speaker could convey more information beyond the text but faced three challenges: style migration, tone simulation, emotional anthropomorphism, especially as the ability to personalize the tone, Can express meaning other than words, such as happy, sad and others feelings. MEITRON can output three independent networks of timbre, rhythm and emotion through spatial modeling of several sentences. He said the network uses unattended training and does not require massive amounts of data to enable dynamic classification.
As for the technical details of Baidu Hongjun's intelligent voice chip, he also provided a detailed explanation on this forum. The chip is suitable for scenarios for speech field interviews, which support the high degree of parallelism between the deep learning calculation and the depth model load.
The core parameters are as follows:
Instruction: HiFi4 custom instruction set, Dual Core DSP
Storage: large memory, including TCM, Cache and SRAM
Process: TSMC 40nm LP, power consumption 100mW
Fulfill vehicle safety standards
Edge data processing capability with remote voice interaction, including group signal processing, voice monitoring, and offline voice recognition.
More techniques such as vision, robots, smart maps etc.
Wu Zhongqin, Head of Baidu's Visual Technology Department and Enhanced Reality Department, released the Visual Semantic Platform 2.0 on the forum. In the 1.0's, our goal is to understand and understand. In the 2.0's, we need to form intelligent interaction, software and hardware.
He said that intelligent interaction includes integrated systems for interaction between humans and computers (faces, gestures, limbs, environment, etc.), which can be applied to AR special cameras and smart car engines, smart speakers and other devices. Considering a soft and tough vision solution, he mentioned PaddleSlim's model compression platform, FaceID's multimodal face recognition component and integrated AI camera solution for hardware and software. It also describes the many benefits of these products.
Yang Ruigang, Baidu's 3D vision scientist, shared Baidu's robotics. He said the robot is an agent. Currently Baidu's research on agent technology focuses mainly on three aspects: autopilot Apollo, construction machines and service robot. He talked about the intelligence of engineering machines that required the agent to move from perception to decision to action. He then explained the technical details needed for these opportunities.
He said "Dr. Tea" in the main forum in the morning is a combined AI capability, including multimotor dialogue understanding, multi-candidate generation, location-sensitive reference elimination, multi-dialogue dialogue status update, and Action triggered. Based on Baidu 3D vision, the robot can perform carpet detection and position tracking. Based on robot pipe planning and control, the robot arm can detect collisions in the work surface and avoid obstacles. After the real-time planning time, the robot can automatically generate the teban. Then the robot can sense the change in the water flow and adjust the position in real time. He said that AI technology will no longer be an island in the future, and the combined AI will promote the development of construction machinery and other applications.
Li Ying, head of the Baidu Maps Division, took a smart estimate of Baidu card transit time based on aircraft. She said that a new generation of artificial intelligence maps combines different techniques such as speech, AR and high quality knowledge maps.
Estimated arrival time (ETA) is the main feature of mapping planning and an important reference factor for users' travel decisions. With this technology, users can avoid overload periods, understand overload times, and avoid overload routes while traveling, but also face many challenges and difficulties, such as major changes, many factors, high timeliness, and strong personalization. Based on this, they proposed a deep neural network computational model, based on the aircraft platform, which can realize real-time ETA accurate estimation and overload of road time calculation. The model combines multi-layer coding and multi-layer sensing.
Finally, Guo Xie, senior manager of Baidu's large data department, introduced the introduction of Baidu's large platform platform. He also interpreted the platform's three core functions, including a secure data fusion management environment and multi-level open customization. Components, as well as a functionally integrated large data development platform.
For more information on Baidu's brain, please click on the address below to learn more.
In this forum, we see, besides seeing Baidu's strong airplanes and data vision and natural language techniques, which cannot be avoided at all AI conferences, the latest in Baidu's automatic machine learning, voice and robotics. progress. In the future, what features will Baidu's brain add and whether the flying paddles will move forward and become the light of the domestic deep learning platform will remain to be seen.
Lei Feng Net Lei Feng Wang Lei(Official number: Lei Feng network)
Lei Feng's original article is prohibited by unauthorized reproduction. See the print instructions for more information.