Mr. Yang Fan: Visual AI Technology Practice

Yang Fan, co-founder and vice president of Shang Tang Technology, member of EGO Beijing Branch. As the general manager of the Shangtang Technology Engineering Products Center, he develops and provides artificial intelligence solutions in the pan-security smart video, mobile internet, and financial industries. With more than ten years of accumulated experience in computer vision algorithmization, product management, project management, R&D management and team management, Yang Fan promoted Shangtang Technology to make significant progress in the commercialization of technology applications. In 2016 alone, it gained hundreds of millions of yuan. Product orders. In 2016, he was elected "Outstanding Young Talent in Beijing" by participating in the establishment of Shang Tang Technology and establishing it as an artificial intelligence leader in a short period of three years.

The following content was compiled by InfoQ's interview with Mr. Yang Fan.

"The greater value of AI lies in combining with different industries."

Yang Fan has been immersed in the field of computer vision technology for many years. During his tenure at Microsoft, he mainly engaged in the incubation of new technologies in computer vision, computer graphics, and other fields, including face recognition, image object recognition, portrait three-dimensional reconstruction; currently the core of Shangtang. The technology is also based on face recognition, intelligent monitoring, image recognition, etc. As the leader of the leading technology, Yang Fan laughed and said that he had laid the hands of the company's researchers. Yang Fan led the engineering development team of more than 200 people to develop and provide artificial intelligence solutions in the pan-security smart video, mobile internet, finance and other industries, and promoted Shang Tang Technology to make significant progress in the application of technology.

Yang Fan believes that AI technology is not new, but it has been intensive in the past two or three years. The key reason is that today we have more informatized processing technologies for voice, images, and video, and we have become stronger in all aspects. Technical reserves. From technology to landing, all of this achieved by AI technology is inseparable from the support of the scene.

AI technology inherits a variety of basic technologies. Solutions for different application scenarios such as industrial, financial, medical, home, autopilot, security, logistics, and agriculture, such as integration of AI and healthcare, should be reflected in smart devices and recognition and diagnosis. The two main aspects: the integration of AI and finance makes financial transactions and management more secure, achieving precision marketing, big data credit, and inclusive finance; the convergence of AI and security enables application scenarios such as intelligent surveillance and security robots; AI, Big Data It doesn't make any sense to wait for these things to talk about the concept. Ultimately, we must return to the scene. Reusable basic technologies and platform tools are important, but only when we fall into the application scenario, we know where the value is.

There has been a criticism in the industry that many companies and developers are actually unclear about the operating principles of deep learning. They only know about applications, but they do not know why.

Yang Fan said: “There are two sets of ideas in the academic world. It is not correct to say that a set of concepts does not know its true meaning. For this concept, Yang Fan recognizes that its realization has already had a lot of teams, including Shang Tang also investing in Conduct more frontier and more fundamental scientific research. “This basic research can guide us to go further in the right direction in the future. However, Yang Fan believes that basic research and applied research must not be neglected. A complete scientific system and continuous directional guidance are very important, but empirical science is also very important. Enterprises must ultimately speak with the results of technology."

The popularity of face recognition has inevitably created a lot of curiosity about the technology and the company behind it. What is the face of Shangtang’s face recognition technology?

For the "Fireface" that has been very hot for the past two years, various kinds of practical scenarios based on face verification capabilities have begun. In terms of Internet information security, account fraud can be better analyzed and investigated, including online mobile phone, desktop, and H5, including customized cameras. The operation logic is very simple. At first, I started to register my face. Now I pay for the face, and the phone gradually unlocks the face. There is also a lot of value in personal certification. The technology of face recognition can determine whether the person who operates the mobile phone is a real person. There is a technical service for living detection, which also includes the form of an offline one machine. Scan the key information of the ID card, including the reading of the internal photo of the ID card and the judgment between the current collectors. Portrait-based authentication is also a very valuable job. It is a special cross-industry Solution. This solution has now begun to spread very broadly from online to offline. For China, the real name system of personal citizenship information is a very important appeal. This appeal can effectively help us solve the security problem of the Internet to a certain extent and solve the public security issues under the line. All online Internet industry applications, to various offline industries, including airports, supermarkets, hotels, will have more and more strong demand for personal identity information verification. Shang Tang also provides a very complete solution in this regard. Program.

Everyone is concerned about the correct recognition rate. In the actual scene, is the correct rate the most critical factor?

In recent years, many companies have invested heavily in face recognition technology and achieved brilliant results. Among them, the recognition rate has always been the focus of various publicity. This year we can frequently see various kinds of reports. 99 %, 99.4%, 99.8% and so on. Although companies declare this way, the difference behind the actual situation is very large, and it will have a lot of influencing factors, so the accuracy rate will be a strong correlation with the industry background and the presumption. It is difficult to make analogy with the recognition accuracy obtained under different scenarios.

When the recognition rate reaches 99%, the difficulty faced by face recognition technology lies mainly in how to deepen this technology in different industry scenarios. Although it looks like 99% recognition rate is already high, different industry scenes have different requirements for recognition rate. 99% may be the entry condition for the technology to be used. In security scenes, the photos are blurred, blocked, and the angle is poor. All bring more realistic challenges to face recognition.

“It seems that homogenization is very strong and simple face recognition. The technical scene of subdivision is actually very complicated. Therefore, it is not meaningful to leave the scene and talk about technology. What can be seen today, including security, mobile phones, etc. Some of the key industries are represented, and there are many challenges to the full-scale deepening of real face recognition technology, which is worthy of us to overcome.”

So how do you judge whether an industry has the value of doing AI scenarios? If you talk about Shangtang itself, what challenges and problems have you encountered in the process of AI platformization?

1, see the demand

First of all, the demand is real. Yang Fan gave a specific example: There is a home appliance manufacturer who wants to use the face recognition function to achieve "I automatically adjust the room to 16 degrees after I enter, and my mother automatically adjusts to 26 degrees into this room." I asked him, "What if you go in with your mother? What if you carry yourself in?" The best solution, he said, was the remote control.

Second, the demand is rigid. Need to consider whether users are willing to pay and how much they are willing to pay? Further deeper logic chains require deeper understanding of the scene.

2, large-scale

It is costly to complete a solution today. Face recognition technology is very different in different scenarios. I do finance today, with 1:1 certification. The error rate is one millionth, one ten-thousandth, and the accuracy is very high. It is very easy to use in financial scenarios. If placed in a security environment, security requires a blacklist of millions of people. Moreover, the blacklist library must be misreported, and there is an alarm for each false alarm. The same is true for face recognition. Different technical indicators and tasks are different under different scenarios. So the same technical concept, the difference in different scenarios is very obvious. Moreover, when technology matures, it needs to have a pre-judgment in the specific demand scenario.

3, data closed loop

To do AI technology, closed-loop data is a very important part. why? When we do video, we find that when your technology is not mature, your business cannot be used, and when the business is not in place, there is no data. If you don't do well, you will end up in an endless loop. How to break this endless loop? The breakthrough in motive power comes from technology, when your technology has a small breakthrough and migrates other scenes. Breakthroughs in technology can lead to the landing of business, the accumulation of data can be brought about by the landing of business, and the accumulation of data can bring about technological advancement. This kind of closed-loop data helps expand the overall business and brings great value. Today, data is confronted with questions and tests of privacy and security. Many technologies including blockchains, as well as non-technical approaches, can lead to deeper exploration.

4, commercialization

It is not enough to make good products, but also to be really valuable in the market, and it can continue to maintain its competitiveness. Any new technology will spread over time, and generally have a time window of up to a year.

During this period of time, how do you view the current situation? How much position does the technology occupy in this scene? Is non-critical application or critical application? Do technological breakthroughs and distributions produce fundamental problems? In the technical barrier period, can we use this period of time to build barriers outside of technology?

Only when barriers are built and the time window is used to turn technological advantages into other competitive barriers is it worthwhile to do such an industry.

5, driven by technological innovation

As early as a year or two ago, we collected a large number of fake photos and videos to attack the face recognition behavior, a variety of cases. When we have a large amount of attack data from real business, we can take very good precautions against all kinds of image video attacks. This comes from the accumulation of business data from a large number of online attacks, and the secondary times of these data. Tap and use.

What kind of revelation does this give us? Doing brushing the face is to do face recognition at first, but later we found that face recognition is not the most important, the most important live recognition, to distinguish whether it is a real person or a counterfeit attack. Only by going into the scene can you discover that the technical challenges you face are different from what you had imagined before. The technical challenges faced by the industry when it is landed actually require redefinition, decomposition and resolution.

From these five closed loops can help us to judge whether the application of an AI technology in a certain scenario is really valuable, whether it really makes sense, and whether it brings greater user value. From these several perspectives, we will have a relatively good conclusion.

Want to do a good job in the landing scene, compound technical talent is particularly important

As Yang Fan said, when it comes to looking at the industry, it is often the use of different technology overlays and combinations. Face recognition and motion recognition are the most critical technologies. However, they actually want to do a good job in landing scenarios. Need a combination of technologies.

Yang Fan said that transforming innovative technologies into actual products is a road full of thorns. It is not easy, and one of the biggest difficulties is how to choose the right direction and timing, and how to find the right talent.

Industry landing requires the integration of various comprehensive key technologies. The needs of the industry are often relatively ambiguous and technically very ambiguous. At this time, someone needs to be able to dismantle them. In Yang Fan’s opinion, finding or developing talents with a background in technology and a deep understanding of the industry is the most critical point for enterprises to achieve AI technology. He said, “Talent problem, team organization issues, development issues, especially the 2B industry, the balance between standardization and non-standards, and the common problems faced by any technical product landing, do AI technology, these issues No one will be less, but it will only be more serious.AI talent is a bigger pit, AI is more technical, and from a past point of view, its combination with the industry is weaker, so you really want to polish a match with the real When the industry needs products, it is necessary to integrate the understanding of the industry with the understanding of the technology. This is one of the most challenging tasks currently, because in the past there may be no such people in the world, and they have an understanding of the industry. Few people."

Conclusion

The landing of visual AI technology and the cultivation of AI talents is a complex and huge topic. It requires a deeper understanding and recognition of technology and talents.

Remote Terminal Unit

Remote Terminal Unit,Power Metering,Power Quality Meter,Electricity Usage Monitor

TRANCHART Electrical and Machinery Co.,LTD , https://www.tranchart-electrical.com