AI application in traditional industries facing three major challenges the premise is the application of high-quality data
Original title: AI application of traditional industry is facing three major challenges the premise is the application of high-quality data recently, artificial intelligence known scholar Andrew Ng published an article, set out his understanding of the application of artificial intelligence slow in the traditional industry.
Whether it is personalized recommendation when the brush a short video, or time-consuming estimated time of take-away delivery, or face recognition when mobile payment to the algorithm as the representative of AI technology is used in consumer Internet industry was "handy." However, traditional industries mentioned, it is difficult to think of a typical case of a very rapid mature application of artificial intelligence. Why AI technology in speed and range far less traditional industries such as consumer Internet industry? AI application of consumer Internet industry more competitive, "it depends primarily on the application of AI technology data, calculate the force and algorithms." Associate Professor Faculty of intelligence and computing Tianjin University Zhu Pengfei introduction, the first data to achieve a certain body mass, which is the basis of the application, in addition Operators must be able to support large-scale force training model, aspects algorithm then need to reach a certain precision, side force should be considered to have some reasoning ability.
Currently only reason why consumer Internet companies in the large-scale application of AI technology, the main consumer Internet companies have more advantages in these three areas.
A few years ago a short video and now so popular, such as the initial development of Taobao, also did not have a strong user stickiness.
With push more accurate sense of the user experience has been greatly improved, the final presentation of subscriber growth spurt.
"Precision push depends on improving the accuracy of the algorithm, and the algorithm to enhance accuracy can not do without vast amounts of data as a basis." Zhu Pengfei explained that in this single scene, the algorithm model needs to evolve, life-long learning.
Since the data is not a closed environment, there is always adding new data, algorithmic models need to constantly be adjusted by learning, iterative upgrade, so that more and more high precision, forming a virtuous circle. "At the same time, although the consumer Internet industry on the accuracy of the algorithm has risen to a certain height, but some traditional industries scenarios, consumer Internet industry are relatively low compared to the AI ??algorithm accuracy acceptable threshold, such as short video, Taobao preference recommendation, Baidu hot search keywords, only need to achieve user-generated sticky, as long as there is a certain accuracy, users can accept.
"Zhu Pengfei said that by contrast, in many traditional industries, the requirements for technical accuracy is much higher.
For example, based on the application of AI technology in the visual aspects of face recognition, high-speed rail stations, airports verify your identity, comparing the accuracy of 1:1% to as high as or even higher can be used.
In the calculation force, the current cloud operator can support large-scale forces have been training 上海哪里有桑拿按摩 model and reasoning, such as short video, Taobao and recommend.
However, a large number of traditional industries in the application scenario, the end of the count on the force intelligent terminal can not meet the requirements of real-time and accurate reasoning. "Compared to electricity providers and social networking systems, closed ecological system makes traditional industries scenarios count the clouds force can not be effectively applied." Zhu Pengfei example, intelligent unmanned systems, for example inspection, electrical inspection, pipeline inspection , traffic patrol, river patrol and inspection requirements photovoltaic mounted on drones and robots calculate the force for real-time inspection requirements, due to the higher complexity of the model video analysis side are often unable to achieve 上海油压 accurate and efficient real-time reasoning, lightweight network while meeting the real-time loss of recognition accuracy.
Since the algorithm can not reach precision requirements, so that the application can not be achieved in many scenarios AI technology.
AI application in traditional industries facing three major challenges Andrew Ng believes in AI applications in other industries outside of the consumer Internet industry are faced with three major challenges: the data set is small; high customization costs; verification process from idea to production deployment very long. In this regard, Zhu Pengfei also emotionally, he analyzed the traditional manufacturing industry as an example.
"The traditional manufacturing enterprises in the manufacturing process to wisdom made the transition, the 爱上海水磨价格 data is a very prominent problem.
"Zhu Pengfei introduction, first of all there is a certain degree of difficulty in access data. Data traditional manufacturing enterprises are closed, because a lot of the traditional enterprise is not new information technology equipment, there is no real-time data collected by the sensors, there is no data center, the data fragmented, missing serious, difficult to obtain consumer Internet companies in the kind of massive, high-quality image data. Next, the data within the industry each plant has a lot of commercial value, so the factories are strictly confidential, which leads to data does not flow, there is no way to share, and then forming a data island effect, affecting the AI ??algorithm optimization model. "we are developing a model when AI algorithms, because the confidentiality of data, often resulting data are to undergo ‘desensitization’, which also seriously affected our judgment.
The traditional industry companies, the lack of technical personnel with AI algorithm development ability of the model, so both sides there are very high barriers to cooperation in the development process.
"Zhu Pengfei said. In addition, the data source is not as traditional industries in the consumer Internet sector as a scene from a single, complex business scenarios result in data tend to be" dirty "and must be" cleaned "to remove a large number of invalid information, AI algorithms to model high efficiency of learning in order to improve accuracy.
"It’s like we teach children knowledge, they talk about knowledge, children can learn quickly, if mixed with a lot of useless information in the knowledge points, children unable to distinguish, learning efficiency will definitely reduce." Zhu Pengfei introduction, and to mark data "knowledge" work is huge and cumbersome, need to have someone to do business, to spend a lot of time and effort. "Traditional manufacturing in order to obtain high-quality data must be production equipment information, intelligence transformation." Zhu Pengfei said that this transformation requires enterprises to invest a lot of time and energy, but also increase the cost of production, which also act as barriers to the application of AI in the traditional manufacturing sector. High-quality data is a prerequisite for application of the past 10 years, most of the development and application of AI is "software-centric" driven.
In support of massive data, continuous optimization algorithms and software, an algorithm to achieve higher accuracy. In the case of the traditional industry can not improve the quality and quantity of data, Andrew Ng believes that traditional industries should be used "data-centric" model, focusing on better quality, higher degree of matching data. "In this line of thinking, the traditional industry has also emerged some good application cases.
AI image identification system such as the medical field, doctors can help to ‘see’ CT imaging films, tumors and other lesions to be identified to assist the doctor’s judgment.
"Zhu Pengfei, due to a lot of data by professional radiologists were marked on the image of the film, so the data more accurate, AI algorithm model made rapid progress in the learning process. At present, many accuracy of the image recognition system can achieve more than 90%, as is the auxiliary doctor, the doctor needs to do the final medical decisions, but this level of accuracy to a large extent reduce the intensity of the work of doctors.
"Although traditional industries have some application of AI technology success stories, but in order to better and AI combined, have to work hard in improving data quality.
"Zhu Pengfei recommendations, first of all for traditional industries have accumulated huge amounts of data, ensuring data security in the premise of open data initiative. Tap the value of the data in the hidden, and associated demand, there will be much room for development. Secondly, emerging industries such as new energy vehicles, while building intelligent plant planning, put the data acquisition, intelligent factors into account.
But Zhu Pengfei stressed, in traditional industries with good AI technology, but also not to abuse AI technology, do the evaluation before the application, if not improve production efficiency, industry as a whole has improved, so blindly forced to use AI technology, is resource waste. "For example, some scenarios require AI algorithms more than 99% accuracy can only be used by evaluating the existing model algorithm can only reach 90% accuracy, then this scene is no need to force launched AI technology.
"" All in all, for the application of AI technology to the data in advance, there is talk about the application of high-quality data, there is no good data is difficult to have a good application.
"Said Zhu Pengfei (Chen Xi). (Editor: Musheng Yu, Zhu Hongxia) share to allow more people to see.