In the age of artificial intelligence, the combination of science and industry became an endurance race. In the history of technology development, the combination of science and industry has always been a difficult issue.Shanghai Science and Technology Co-founder and CEO Xu Li It seems that in the era of artificial intelligence, the combination of the two has become an endurance race, and This task is becoming more urgent. For Shang Shang Technology, which was established in 2014, the team has always had a rich academic atmosphere and the company has close cooperation with academic teams at home and abroad.But the intersection of science and industry must be two-wheel-driven, in the process of policy, Capital, technology, floor space, scene applications and other factors are indispensable.Shang Tang from the perspective of science and technology, they are more concerned about computer vision, artificial intelligence, this part of the eye, hoping to help people see and understand the world. There are many forms of landing in the field of computational vision that bring a perceived breakthrough: for example, the camera is perfectly imaged when the environment is very dark, and the AI can help the expert to read the map, but before that, it must be clear By the people to get through the whole process. Xu Li believes that from Shang Tang's attempt, artificial intelligence now happens to come to a very good time, external elements are accelerating the acceleration of AI landing. The following is a speech by Xu Li at the Innovation Conference 2018 in Geek Park (edited by Geek Park) Science and industry: from contradictions to convergence Hello everyone, I am Xu Li of Shang Tang Technology. We all know that this wave of artificial intelligence certainly has a very large relationship with many scientists' undertakings. In a sense, the science industry is now a crook, but this is only what happened in recent years. I think When I was in school, I had always had such a passage. We were going to read a Ph.D. Many professors at overseas elite schools said that in fact, reading a doctorate is not a path to getting rich. In general, it is certainly necessary to have more money to read a doctor and be a scientist. Rather than being a scientist, for the sake of having money, the two logics are counter-productive. In fact, in the Middle Ages, basically all argue that the nobility can do real scientific research because the contradictions and conflicts between science and industry are very big and the two are totally different points. Why is this stage now, It seems that these two lines have come to an intersection, and we can actually look at the obvious changes in the industry. One big difference between science and industry is that we define scientists: First of all, everybody earnestly studies scientifically and focuses on technological breakthroughs and breakthroughs in basic theories. In fact, many times, even 100 years ago, many theories were innovated and developed. However, in the true sense, There is a combination of very few, all of us are familiar with a variety of scientists, but in the true sense of scientific success, and the final industry are far apart, the industry concerned about what it is concerned about the real profitability, standardization, rapid realization , Solutions, and that the real industry actually comes from the market's demand-driven. Therefore, science and industry is still a completely different concept of such a concept. Since the difference between the two is so far, what do we think of this matter under such circumstances? What exactly happened in the current stage so that science and industry can meet in one place? I think it is crucial The point is that the intersection of science and industry must be two-wheel drive, what is the two-wheel drive? First, the technological breakthroughs in the core are faster, and the development of science will have so many years and science and technology develop steadily, that is to say, technological breakthroughs are incremental. Theoretical studies are very important in this process, but these studies may or may not Can be really applied to all we think, cognitive business up. In turn, science and technology are sometimes leaps and bounds, it is not a linear, incremental .For example, this wave of artificial intelligence breakthrough, in fact, is a sense, the traditional manual guidance of intelligence, into Pure data-driven intelligence, including deep learning that we all know nowadays, pure data-driven approaches based on deep learning, is actually a subversive change, not an incremental change. Only in this state Can we bring about entirely different new productive forces. In the other direction, let's talk about industries and industries. What kind of conditions do we need? We can also see that the subversive technological development I just mentioned has actually happened in history. For instance, some new materials, new energy sources Found that, in fact, have, but this may not really be equivalent to the fact that it can be realized in the industry, to change. For example, when we talk about tungsten, the material has been discovered for a long time and really used in light bulbs. As a lamp, it still needs to be pushed back from a realistic scene to material breakthrough and material revolution. This is a Very important point. What is important from an industry point? The important thing in the industry is that the needs of the industry really come to a critical point of eruption, that is to say, the general public, we are actually well educated. In the meantime, advances in science and technology can bring about subversive workforce and productivity. In addition, the market is educated to bring about what we are all aware of. With these two things, we feel that science and technology Promote and industry changes, together, leading the progress of mankind, changing human life, this is very important. Shang Tang technology attempt From our perspective Shang Shang focuses on this part of computer vision and artificial intelligence, and we hope to help people see and understand the world. From this perspective, what are scientists doing in the future? , We are aware that the essence of the core is to say that I have a better ability to perceive, and the ability to perceive not only from the hardware point of view, algorithm, theory, and behind Support, in fact, brought the concept of core breakthrough in perception. Now most of the sensors actually handle the visible light we think. As you can see, this year's breakthrough in the visual field has spread from visible light to invisible light. We can have a UV camera that can have infrared Camera, far infrared, near infrared camera, and even a variety of thermal cameras, these cameras open is a completely different area of application. For example, iPhone X now supports face unlocking or paying, in fact, because our perception has shifted from visible light to near-IR, which we think of. In the near-IR scenario, we can do better live-action recognition , Better verification, so give us a real pay, or online and offline scenes played a role in escorting. In addition, we can see that there are many cameras on the road. At night, many people think that the camera can not see the camera, but the perception of the camera has now reached the level of stars, and even Aurora, what concept? In the case of very dark , It can form a very perfect imaging, the camera on the ground so, the sky camera as well. Shang soup has always believed that, for the moment, the development of the industry is the application of the ground camera, but in the future, more and more cases of the sky camera, we all face the same problem of data processing, so in remote sensing imaging, very Already used the super-sense of super power. For example, remote sensing images can have a variety of different spectra, can detect clouds, snow, distinguish plots, in doing a wide range of different areas, you can achieve different processing, the different scenes of the data connection and splicing. There's a core boost in basic perception that is an extension of the visible range. In addition, our real-world scene is 3D stereo, and the picture is always 2D, and 2D to 2.5D to 3D extension, it will bring very different changes in our application. In fact, we have always believed that this perception of the scene brings new interactive capabilities, and interactive ability to change has always been the beginning of the Internet revolution.I do not know if you remember the iPhone sliding interaction just came out when there is a game very Hot Angry Birds is because of its design philosophy, which is very much in line with the human-computer interaction mode of a mobile phone. However, when our sensing system perceives 3D better, the object in the 3D scene can be better presented It is possible that our next new interaction model will emerge based on new 3D perception and AR / VR / MR technologies. We can see these three scenes, the left is a simple use of SLAM technology, in perception of 3D cases with some different scene content in the middle is that we are in the office, flat ground has made a small animal and a tree As you can see, this rendering mode will be more and more realistic, from the original 2D photos, has been highlighted in the application of the entire 3D which is the use of a normal camera. The right is that we put some real things on the table, two little monsters are fake, if we look behind the chairs, butterflies are all fake. This application brings a lot of different experience . Imagine if the future of our single camera phone, we can bring some different 3D experience, I think 2018, 2019 will produce more and more game scenes, social scenes, is based on And the reality of a more complete integration, which is what we are talking about the concept of scientists to do is to enhance the perceived ability. On a theoretical basis, when scientists make big data-driven artificial intelligence, we find that there are many inherent deficiencies. For now, there is a lack of solvable things. In the past, if the artificial guidance of intelligence, what is the concept, we use big data for statistical learning, statistical learning of human beings will add a lot of prior knowledge, we consider the age distribution of people are evenly distributed or normal distribution, these I have added a priori distribution. I added a priori, a certain significance for the prediction is a great reference. However, if the data is getting bigger and bigger, maybe I do not make a priori assumptions about the data. In this case, a problem arises. All the laws are learned from the data and will produce a Big flaw, called lack of interpretability. The other big data problem that we really want to solve is that we might say that we need to give more data-aware and machine-aware labels, and the so-called labels give some correct answers, but not Everything has the correct answer, so we still lack supervisory information. In some specific situations, such as medical treatment, we have seen a hospital research institute. He said that there are two national experts on our side who also hope to bring knowledge to the real church Computer learning, but just because these two veterans can see the early illness on the film, no one in China would do it and asked if we could do it. In fact, the definition of the machine to see things clearly, the machine can do is define the output input is very clear, I give a film, it can indicate that this is not a certain early patients, this thing is the definition is very clear Unfortunately, there are too few samples and not enough samples. If we need to do such training, we usually need a million data on how many old experts can read a film in a day. Hong Kong doctors can not read more than eight films a day and more than eight are entitled to be read as weary newspapers. According to the concept of reading 10 sheets a day, veteran experts need to carry out 300-500 years of diagnosis to be able to provide enough data, and no matter how old the experts are already old, even if they were veterans from an early age In fact, I think it is very difficult, that is to say, there is no such new sample, but also a variety of interdisciplinary and multimodal interaction changes. Therefore, after these, in fact, what we are doing nowadays scientists are basing his theoretical foundations into the original research questions, such as multimodal, multi-task learning, weak supervision and Unsupervised learning even says that in the absence of a sample, I migrate samples and even generate sample learning. There is another way I can get through this loop, although I do not know if this is a good answer, but I can do to enhance our learning to help us get through. For example, if you look at the description of the problem, the bird has a white chest, a light gray head, black wings and a tail, and you will not feel particularly surprised, but in fact the process is counter-productive, We gave a description of the Chinese, he produced such a photo, which means that there is currently no bird in the world like any other bird on this photo, that is to say, we can get through the real meaning , Generate sample data from the description of the text, sample case. Going forward, if you can create a picture with text, is it possible to generate the video in text, the director may not have to work, I can write a book to generate video.Science is gradually put some originally thought of any impossible Things become possible. Can see that with such a technology, it is possible that we will never trust the circle of friends again after the forwarding of the content, everything is possible .But these breakthroughs in science and technology, from the perception of ability, theoretical basis for the promotion, in fact, occurred The change and the real industry is still a long distance. We look at the industry in the end what did it do? Industry AI breakthrough, the first is that the cloud + end of the model to get through, the cloud is the server, computing power to enhance, and the end is from a variety of devices, there are more and more Computing capacity, in order to meet the AI technology breakthrough, in fact, more and more equipment has been prepared smart, or be intelligent, is to see, there is no suitable algorithm into my car inside the robot, the phone Inside, and even on a variety of handheld devices, this is a very interesting change that has taken place in the industry. From our point of view, AI technology trends are multi-scene, multidimensional linkage, all-stack innovation, from the bottom of the algorithm to the middle of a wide range of technology modules, to the above core applications, this part of the core application In fact, in fact, from the industry to a real demand, which the needs and requirements are not the same. Because in many cases, we also deal with a lot of enterprises, companies say that we need to be AI.This is a proposition, found a bunch of problems, some of these problems we are not doing well, is not using AI Method to adjust to you, I want to say, impossible, hard to do. What AI can do? Just now I said, when a problem definition, input, output, clear, and someone to get through this process, use AI to replace these abilities, will be more natural. So what AI is doing is to improve its productivity in application scenarios. This is something that can be done, but if people do not know how to do that, you have to let AI help you sort it out and deduce it , Which is very difficult, so I think the role of this scene can sort out the real needs to the industry. How to Promote the Integration of Science and Industry With the convergence of industry and science, we think there are two big pieces of external elements: Elements of a temperature environment. Element two, from the perspective of industrialization, we talked about three. What is the external environment? Now I think that external management actually played a very good role in accelerating and catalyzing the role of China. It is the first time in history that China has promoted the development of the entire AI from the national level. In fact, you can see Most countries in Europe and the United States actually act as suggestions since the beginning of the year. However, since China has made the entire move forward, the warming of its policies, including the formulation of some white papers, is often from the perspective of the people as a whole and the whole nation. Help industry fall. The second is capital increase, and we can also see that at this stage, the capital for artificial intelligence, such a hug of technology, in fact, made the industry from the original core technology breakthrough, to the industrial chain which shortened a lot. The third key point is that the resources are available. If you want to react nuclearly, you still have to have oil. The so-called complete resources is that, at present, the computing power reached a certain level of computing resources, and even based on the above platform, and some operating system research, have reached a point, so that the external environment is already allowable Science to the industry for a very good conversion. From the industry itself, it is necessary to three: First, technological product. It is very hard for an enterprise without an AI to be realized by actually selling technology. A breakthrough in a core technology, as I just said, if we say that we have nothing to generate this bird, but we need to turn it into a commercial realization , We must have a corresponding product to do the carrier, this step is actually most companies gradually completed, technology, technology, we can become a kind of precise identification ability, or a chip, cloud and client These two modes are the core breakthrough in product technology. Second, large-scale landing. If we say that we need to truly form a technology of a large enough influence, we need it to have a large-scale ability to cover in one city and two cities is useless, we need to have a rapid growth capability, Or, these things can be standardized into a product which, then its promotion will be very successful. In fact, many people use the device, which has generated a lot of AI, just as we used to take pictures with a cell phone, now an ordinary camera can achieve the look of the SLR, in fact, this is an AI technology, landing to Centralized expression of mobile phone products. Thirdly, it is not clear in which industry the AI is able to form a dramatic change in the industry. Therefore, scene diversification can in some sense create a good new technology among technologies. We can see that there will be such a change from the public service, the personal application and the entire social management. In fact, it is a bit like the technical breakthrough when we talk about the first industrial revolution. We have a good steam engine, However, if there is no diversified scenario application, in fact, the technology is actually realized and the industrial revolution is really promoted. In fact, there is still a long way to go. At that time, the application was: First, I can do large-scale industrial manufacturing. Second, I can harvest and irrigate agriculture. Third, I can even do railways and transportation. At that time the world's largest market value of the company is to do the railway transport. So from this point of view, the diversification of the technology scene, but also this wave of AI core elements of the floor. For example, personal application, the phone will have some such scenarios, the right is the help we introduced pictures to smog Mobile phone application, that is to say your camera, if you take a haze day in Beijing, do not worry about a key to eliminate hazy days. In the middle is the management of some photo albums. I believe now that the pictures taken by everyone are already manageable according to people. My own photos, my friends, children and family can be divided into different groups and the future can be based on more dimensions and Tags for segmentation, for example, according to the age and intimacy, the machine can help you through the automatic classification, combing. The leftmost photo, we are talking about let AI have more creative elements, in fact, we can simulate a lot of artistic effects in the video, you can even learn Van Gogh, learning Monet, you can see the future of mobile phones Shooting a photo, some time ago we are imitating Qi Baishi shrimp, in fact, I think do not imitate, you take a really eat shrimp with an algorithm, you can immediately become Qi Baishi style. So this AI and personal application will bring a lot of entertaining process.Another angle is the AI and the governance of the community as a whole, will also play a very large depth of integration. We can imagine that in the current state, the governance of the whole society is still detached. Even if the use of technology is not taken as a whole, we think of all aspects of the basic necessities of life, whether from the ground to the sky or from every aspect of the ground, Walkthrough and evolution, you can do a holistic plan. For example, we travel, on the one hand in-car AI can help unmanned, obstacle avoidance, better navigation services on the other hand in the sky AI, technology through real-time perception of the ground situation, and even The roads are well planned, and the connection between heaven and earth achieves better synergy. There may be many things that can change better in the not-too-distant future, which is where we see a big connection between technology breakthroughs and scenarios. The ultimate AI is not the product, is the public service, the future, because the AI can do things in front of predecessors, like things on the counter, we provide a variety of certification on the counter, people are as a Assisted docking of the mouth, and then I believe that the docking of the mouth, will be replaced by the AI, then we provide services in the future, in a real sense can be done according to the different state of each person gives different customization Service Speaking so much, I say science and industry, or that there will be a long way to connect these two things in the middle, so we say that AI is an endurance race even from our AI business. Science and industry can eventually be combined at one location and for the moment we are at a very good time because the surrounding environment is catching up and accelerating it. So we also hope that at this stage, AI can really help everyone to make life better for everyone. Thank you.
|
