In order to realize the application of AI technology, firstly, the data must reach a certain volume. In addition, the computing power must be able to support large-scale model training, and then the algorithm needs to reach a certain accuracy, and the end-to-end computing power must also have a certain reasoning ability. At present, only consumer Internet companies are applying AI algorithm technology on a large scale, mainly because consumer Internet companies have more advantages in these three aspects. - Zhu Pengfei, associate professor, Department of intelligence and computing, Tianjin University Recently, Wu Enda, a well-known AI scholar, published an article on his understanding of the slow application of AI in traditional industries. Whether it is personalized recommendation when brushing short videos, time-consuming estimation during takeout delivery, or face recognition during mobile payment, AI technology represented by algorithms is "handy" in the consumer Internet industry. However, when it comes to traditional industries, it is difficult for people to quickly think of the typical cases of very mature application of artificial intelligence. Why is the application speed and scope of AI technology in traditional industries far lower than that in consumer Internet and other industries? AI has more advantages in consumer Internet industry "The application of AI technology mainly depends on data, computing power and algorithm." Zhu Pengfei, associate professor of intelligence and computing Department of Tianjin University, introduced that first, the data should reach a certain volume, which is the basis of application. In addition, the computing power should also be able to support large-scale model training, and then the algorithm needs to achieve a certain accuracy, and the end-to-end computing power should also have a certain reasoning ability. At present, only consumer Internet companies are applying AI technology on a large scale, mainly because consumer Internet companies have more advantages in these three aspects. A few years ago, short videos were not as popular as they are now. For example, Taobao in the early stage of development did not have strong user stickiness. As the push becomes more and more accurate, the user's sense of experience has also been greatly improved, and finally there is a blowout of user growth. "Accurate push mainly depends on the improvement of algorithm accuracy, and the improvement of algorithm accuracy is inseparable from the massive data as the basis." Zhu Pengfei explained that in this single scenario, the algorithm model needs continuous evolution and lifelong learning. Because it is not a closed data environment, new data will always be added. The algorithm model needs to be adjusted and iteratively upgraded through learning to make its accuracy higher and higher, forming a virtuous circle. "At the same time, although the algorithm accuracy of the consumer Internet industry has risen to a certain height, compared with the application scenarios of some traditional industries, the threshold accepted by the consumer Internet industry for AI algorithm accuracy is relatively low. For example, short video, Taobao preference recommendation and Baidu hot search keywords only need to achieve the purpose of user stickiness as long as they are accurate Users can accept it. "Zhu Pengfei said that in contrast, in many traditional industries, the requirements for technical accuracy are much higher. For example, the application of vision based AI technology in face recognition can be applied only when the 1:1 comparison accuracy is as high as 99.99% or even higher for identity verification at high-speed railway stations and airports. In terms of computing power, cloud computing power can support large-scale model training and reasoning, such as short video, Taobao recommendation, etc. However, in a large number of traditional industry application scenarios, the end-to-end computing power on the intelligent terminal can not meet the requirements of real-time and accuracy of reasoning. "Compared with social networks and e-commerce systems, the closed ecosystem of traditional industry application scenarios makes cloud computing unable to be effectively applied." Zhu Pengfei, for example, takes intelligent unmanned system Patrol inspection as an example. Power Patrol inspection, pipeline patrol inspection, traffic patrol inspection, river patrol inspection and photovoltaic patrol inspection require that the computing power carried on UAVs and robots meet the requirements of real-time patrol inspection, Due to the high model complexity of video analysis, the end side often can not achieve accurate and efficient real-time reasoning. The lightweight network not only meets the real-time performance, but also loses the recognition accuracy. Because the accuracy of the algorithm can not meet the use requirements, the application of AI technology can not be realized in many scenes. Application of AI in traditional industries faces three challenges Wu Enda believes that in terms of AI applications, industries other than the consumer Internet industry are facing three major challenges: the data set is very small; Customization costs are high; The process from validating ideas to deploying production is long. Zhu Pengfei was also deeply touched by this. He analyzed the traditional manufacturing industry as an example. "Data is a very prominent problem in the transformation of traditional manufacturing enterprises from manufacturing to intelligent manufacturing." Zhu Pengfei introduced that first of all, there are some difficulties in obtaining data. The data of traditional manufacturing enterprises is closed, because many traditional enterprises are not new information equipment, there are no sensors to collect real-time data, and there is no data center. Therefore, the data is scattered and seriously missing, so it is difficult to obtain the massive and high-quality data like those in consumer Internet enterprises. Secondly, many data of factories within the industry have commercial value, so the factories are strictly confidential, which leads to the non circulation of data and no way to share, thus forming the data island effect and affecting the optimization of AI algorithm model. "When we develop an AI algorithm model, because of the confidentiality of the data, we often get the data through 'desensitization', which also seriously affects our judgment. However, enterprises in traditional industries lack technicians with AI algorithm model development ability, so there are high barriers in the process of cooperative research and development between the two sides," Zhu Pengfei said. In addition, the data sources in traditional industries do not come from a single scenario as in the field of consumer Internet. Complex business scenarios often lead to "dirty" data, which must be "cleaned" to remove a large amount of invalid information, so that AI algorithm model can learn efficiently to improve accuracy. "It's like we teach children knowledge. Only by talking about knowledge points, children can learn quickly. If a lot of useless information is mixed in the knowledge points, children can't distinguish it, and the learning efficiency will be reduced." Zhu Pengfei introduced that the work of marking "knowledge points" to data is huge and cumbersome, which requires special personnel in enterprises to do, and it takes a lot of time and energy. "If the traditional manufacturing industry wants to obtain high-quality data, it must carry out information and intelligent transformation of production equipment." Zhu Pengfei said that this transformation requires enterprises to invest a lot of time and energy, and will increase production costs, which has also become a barrier to the application of AI in the traditional manufacturing industry. High quality data is the premise of application Over the past 10 years, most ai r & D and applications have been driven by "software centric". With the support of massive data, software and algorithms are continuously optimized to obtain higher algorithm accuracy. In the case that traditional industries cannot improve data quality and quantity, Wu Enda believes that traditional industries should adopt the "data-centric" model and focus on obtaining data with better quality and higher matching. "Under this idea, some good application cases have also emerged in traditional industries. For example, the image recognition AI system in the medical field can help doctors' see 'CT images, identify tumors and other lesions, and assist doctors in making judgments." Zhu Pengfei said that because a lot of data are marked on the images by professional radiologists, Therefore, the data are relatively accurate, and the AI algorithm model has made rapid progress in the process of learning. At present, the accuracy of many image recognition systems can reach more than 90%. Because they are auxiliary doctors, doctors still need to make medical decisions, but the accuracy of this level reduces the work intensity of doctors to a great extent. "Although traditional industries have some successful cases of applying AI technology, they still have to work hard to improve data quality in order to better integrate with AI." Zhu Pengfei suggested that first, for traditional industries that have accumulated massive data, they should take the initiative to open data on the premise of ensuring data security. Mining the value contained in the data, associated with the demand, will have a lot of room for development. Secondly, for emerging industries, such as new energy vehicles, the factors of data acquisition and intelligence are taken into account when building intelligent factory planning. However, Zhu Pengfei stressed that while making good use of AI technology in traditional industries, we should not abuse AI technology. We should make a good assessment before application. If we can not improve production efficiency and improve the industry as a whole, then blindly and forcibly using AI technology is a waste of resources. "For example, some application scenarios require AI algorithms to achieve an accuracy of more than 99%. Through evaluation, the existing model algorithms can only achieve an accuracy of 90%, so there is no need to forcibly launch AI technology in this scenario." "In a word, for the application of AI technology, we should put data first and talk about application with high-quality data. It is difficult to have good application without good data," Zhu Pengfei said. (Xinhua News Agency)