Netease Technology News December 18 news, at today’s NetEase economist annual meeting artificial intelligence forum, despise technology chief scientist Sun Jian, KeTV Tuo founder Shan Shiguang, the fourth paradigm founder and CEO Dai Wenyuan, Tong Wei, the vice president of Shield Technology, participated in the first roundtable dialogue. The dialogue was hosted by Zhang Xiaodong, Chairman of Wuzhen Zhizhan and author of "A Brief History of Artificial Intelligence". The theme was "AI technological change: trends and directions."
Sun Jian admits that there are certainly misunderstandings about AI in the public at the moment. Through the media, events like AlphaGo have made everyone feel that artificial intelligence has arrived soon, but in the academic world, people actually think that artificial intelligence is still very In the long run, there are still many basic problems that have not been resolved. Shan Shiguang strongly agrees with this view. He also likened this misunderstanding to a story like a blind man. "Everyone's experience is based on his own past research experience or development experience, so even practitioners of AI Understanding is also one-sided, like the blind people's feelings. Everyone does not have a general idea, because no one can make all things very clear at such a time."
In Dai Wenyuan's view, artificial intelligence and human intelligence are two different things. The working principle of the human brain is different from the working principle of the machine brain. For example, there is a brain with particularly high computing ability and extremely fast memory, but it It is not the same as the person's induction, abstraction, and summary.
Gu Wei summed up this misunderstanding as "the truth will always have a process." "The future of artificial intelligence is just like the Internet. At the time of the Internet in the late 1990s, everyone felt that it was also questioned and misunderstood it. Now there are giants such as BAT. I believe that in the next 5-10 years, we will not rule out the existence of future artificial intelligence BAT." Gu Wei said. (Zion)
The following is the main content of this roundtable dialogue, which is summarized by Netease Technology:
First, is there any misunderstanding in the public's understanding of the concept of artificial intelligence?
Zhang Xiaodong: When I was studying in the United States thirty years ago, three groups of people and a group of people in the computer were theoretical. There were also groups of people who were systematically involved, including software and hardware. There is another group of people who do AI, systems and theory people look down on each other at the time, but they also look down on people who do AI. This is 30 years ago, and the situation is completely different now. Today I would like to ask several guests on the stage. Are there any misunderstandings about AI?
Sun Jian: Certainly there are some misunderstandings, because through the media, events like AlphaGo have made everyone feel that artificial intelligence has arrived soon, or that many things can be done. People in academic circles think that artificial intelligence is still very long-term. There are still many basic problems that have not been solved, but I think this is quite normal.
Shan Shiguang: I also agree with Sun Jian's point of view. It must be misunderstood, and these errors can be used as an analogy. For example, how do you understand the story of the blind? There are indeed many experts, scientists, and technicians in the community who are doing artificial intelligence, but in fact, most people, or nearly 100% of them, are doing different aspects of AI. For example, they may not be too visually too much. To focus on listening, listening is basically not so concerned with natural language understanding. Now this gap can definitely give everyone more communication, but each person's experience is based on his own past experiences of some research or development experience, so even if the practitioner's understanding of AI is one-sided, like the blind people feel like, everyone There is no overall idea, because no one can make all things very clear at such a time. It may be over-generalized, such as generalizing AlphaGo's technology to other problems. Computer vision people will also unconsciously generalize their progress in this area to other areas. Will lead to more misunderstandings.
Dai Wenyuan: I think this misunderstanding is often caused by overestimation or underestimation. Let me give you an example. When I first started my business, it was 2015. At that time there was no AlphaGo. I talked to a lot of people and said "artificial intelligence". After finishing the four words, they said that I was not interested in artificial intelligence. We excessively use Baidu artificial intelligence technology to help the company provide 8 times the income, he said you this is flicker, right? is that a lie? It's not the same today. I haven't gone yet. They may be interested in artificial intelligence. When I said that it was increased by 8 times, he said that this is not enough, but even more. Before we were overly pessimistic about this matter, we now feel over-optimistic. It feels like artificial intelligence can go far beyond people in all aspects. Even the most asked question is whether I will lose my job. This is on the one hand. On the other hand, I think there is a big misunderstanding, that is, people will establish a connection between artificial intelligence and human intelligence. We think that the future of artificial intelligence is such a capability that will meet or exceed human intelligence. But in fact, in our opinion, artificial intelligence and human intelligence are two different things. I will give you another example. People will feel more like the way people know a person and how a dog knows a person. Is not the same, people's ability to know people is strong or dog's ability to know people? In fact, not necessarily, the dog's ability to meet people is stronger. The same reasoning is that the working principle of the human brain and the working principle of the machine brain are not the same, that is because these two are different species, and each species works in a way that he is better at. What special features are we using today when we are doing artificial intelligence? There is a brain. This brain has particularly strong computing power. When it is very fast, how does it generate intelligence? It is different from people's induction, abstraction, and summing up the actual methods and principles. It is impossible to expand in detail today. If we have the opportunity, we can actually discuss this topic in more detail.
Gu Wei: When it comes to retreat, indeed, because everyone is talking about artificial intelligence, and even when we met with the founder Jiang Yu of the shield technology last time, the host also called him, saying that it is now personal to talk about artificial Smart, so Chiang Kai-shek said that I am embarrassed to talk about the artificial intelligence that we are doing real financial technology. But feeling is still a trend after all. I think the truth will always have a process. The future of artificial intelligence is just like the Internet. At the end of the 1990s, people thought that it was also questionable about it and misunderstood it. To this day, some giants such as BAT have emerged. I believe that in the next 5-10 years, there will be no BAT for future artificial intelligence in the companies that are present here. Thank you.
Second, is the dynamic of artificial intelligence's recent revolution really an algorithm?
Zhang Xiaodong: The next question. I want to tease about “the masses fight against the masses.†Last week, Zhou Zhihua, NJ’s Chou and the Computor’s High Teacher, both had an interaction on Weibo. What caused the AI ​​revolution in recent years? Most of the cognitions are artificial intelligence revolutions because of the increase in power and data. Professor Zhou expressed different opinions. He said that the fact that the algorithm has actually improved in the past few years has also made a tremendous contribution to artificial intelligence. I would like to hear several guests on the stage. In order to express their firm position, you can only choose one position. Which one do you choose?
Sun Jian: My position is the same as that of Chou, because in fact, in terms of computing power and big data, we look at many of the issues that we study. In fact, we can study small data and when the computing power is not very large. The algorithm is not a point, but there are many, many points in this algorithm. The important points may add up to more than 10, and these points add up to let ten years ago, twenty years ago, deep learning, or artificial intelligence network The training may be presented by some top masters at the time and other teachers may not be able to present it. Today, through many efforts, especially improvements in various optimization algorithms, so that the vast majority of data can be presented, so that this thing can be done. Today there are still many small data in the study of algorithm improvements, how to test large data? I think the algorithm is more fundamental.
Shan Shiguang: I don't know where the next two are. I'm standing opposite Mr. Sun. More people will find the algorithm very important. I also thought that the algorithm was very important at the time, but I wanted to stand on the opposite side of the table so that I could have more perspectives. In the mid to late 1980s, artificial neural networks of that era, including multi-layer neural networks, and now in fact this round brought the largest performance-enhancing convolutional nerves in these areas of computer vision, including speech recognition and so on. The network was born in that era. The Convolutional Neural Network was formally named in an article by Professor Yann Lecun in 1998, but in fact it was earlier. In 1989, it was used to identify numbers. If we go back further, it was in 1980 that the Japanese scholar Kunihiko Fukushima's approach to understanding neural networks such as neuroscience was essentially the same in practice, and it had been proved in the late 1980s. Layer neural network, the so-called depth is actually the number of layers, multi-layer neural network is sufficient to theoretically approach all the functions considered to be complex.
However, the large amount of work we are doing now is a problem of approaching complex nonlinear functions like this. Therefore, if we say this, in fact, in the 1980s and 1990s, the law of defamation was very much like the present. By 2012, more layers will be added, but there are indeed many more improvements and more. However, we recall that if there was not a large amount of data in the late 1980s, there would be no current calculations. If it is at that time that there is current calculation power, I think this progress may have occurred at that time. . Including big data, algorithms, computing power. Why is it impossible in the 1980s? At that time, if the computational power of the current calculation model is estimated to be at least one year, because it is now going to do deep learning, coupled with the GPU server, many things nowadays have to train for as little as three or four days. It takes a month. If the computing power of that era were to be such a training, it would be really impossible for a year. The development of an algorithm, the emergence of a technology, ran up a year ago, and only after one year can we see the results. The result is good. If there is no problem, then it will be impossible to re-adjust it. This is impossible. Therefore, I think that the improvement of the force is one. It is not as good as it is now. It owes the east wind and owes gasoline, if it is to be ignited. Dongfeng is a high-performance computer, and gasoline is big data. So I think that this round of progress really is that people who do algorithms really want to thank those who do the system, and also thanks to the emergence of new infrastructures such as the Internet and the Internet of Things, which made us have big data.
Dai Wenyuan: My point is that although these are important, what actually matters most in the moment? Often the places that are not so much attention are the most important at the moment. For example, data, data, I think is important, but in fact I am most concerned about the data is ten years ago, because do AI without data for ten years ago, we can not do it, then came to the Internet, with data, 2008, 2009 Concerned about the power of calculation, companies like BAT have money and can buy machines, and they can't do BAT. To BAT to do the algorithm, what algorithm? Actually, it is not a problem of the past algorithms, but the past algorithms can not increase the effect with the increase of the amount of data. To be an algorithm that will become better and better with the increase in the amount of data, in fact, these are actually everyone. Full attention. One of our big feelings is that, for example, we are now investing a lot of energy in the architecture. The architecture is a problem that is often overlooked by everyone now. Many people think that doing AI is an algorithm. After the algorithm is done, Find a programmer to implement it on the line.
But in fact, the difference is very large, because this involves too professional problems. Let me give an example. Let's start a company. Does a person control 100,000 people for the same management method? The method of managing one person, ten people, one hundred people, and ten thousand people is not the same. It is different from the same machine, ten machines, one hundred machines, and ten thousand machines. When we started our business, we didn't have any management level in our houses. Any two people could explain each other. But when the company reaches a large number of people, it can let any two people talk to each other and it is finished. At the same time, there are 100 people looking for it. When one speaks, that person collapses. So why is there a reporting system and hierarchy? It is because of the need for a better mechanism. In fact, computers are the same. Computers do not scale horizontally. That is a good idea for us. It is not a real horizontal expansion. It requires a good architecture. This architecture can support you to put so much power into it. It is at this stage that I think the entire performance has been neglected, and many and many systems, many and many applications have not been able to be made at last. It is because we are doing too badly in this place.
Gu Wei: My position is, in fact, I agree that data and calculation are the main factors. From the aspect of calculation, in fact, Sun Quan has just talked about it. Let me give you an example. He said that because of such a large amount of data, we The algorithm, which was implemented at the time of many algorithms, also existed many years ago. It was developed in the model. When these algorithms were run and these models were run, the time was very long and could not run at all. This is one aspect. Another aspect is that there is a process of model deployment. That is the process of model development. Development will take a long time. You also have a process of model deployment. Let's take an example. In the past, for example, we had a model in FICO in the United States. It was a swiping card and credit card transaction. It was actually a neural network model. At that time, we did a lot of work to ensure its real-time calculation. At that time, the promotion cost was very high, so many small ones could not get on. Therefore, only one large cloud service could be provided, and the larger ones would go up. Our China is basically blank in this section. Now it is different. Because there is a computing power here, we have developed it in a relatively short period of time. Our deployment also meets the real-time requirements. This is an increase in computing power.
The second is data. Many of them are now online. They include online financial applications. It is very easy to apply for a loan with a mobile phone. At the time of the mobile phone's submission, including the corresponding actions of the mobile phone. Data, with this data, we can do more of the overall model, not the previous tradition may be to look at the previous information, I do some verification. With more data and the improvement of these computational powers, new algorithms and new models have also been applied. Therefore, Zhang Zhang, my opinion is still standing on the data and computing power.
Third, there is no financial risk? Does a large amount of capital injection mean bubbles?
Zhang Xiaodong: If Zhou Zhou can come, the majority still stands in his ranks. We have experienced two ups and downs of artificial intelligence. This time in artificial intelligence, first of all, from the scale of financing, there are several companies that have absorbed large amounts of money. The last time we talked at Microsoft’s committee, Shen Xiangyang asked the same questions. First, is there a problem with funding risks? Second, does a lot of capital injection mean bubble? This, I would like to ask Mr. Sun to speak first, and less money can be challenged.
Sun Jian: The first two artificial intelligence bubbles, this time is not a bubble, the biggest difference is that this commercialization process is very solid and compared with the previous two. We despise that science and technology have invested a lot of money not because of the deeper strategic layout and in-depth investment development. Computational power is very important and depends on greater computing power. The biggest difference is that the business is doing very well. This money is mainly used for strategic development, so I think this time is really different from before.
Zhang Xiaodong: I slightly interjected, because in fact it is not the industry of artificial intelligence. In the entire environment of China, we actually experienced several times, for example, the industry that drove two years ago, recently shared the bicycle industry, and started. There are also large-scale financial interventions, and mergers will soon follow because of their homogenous competition. So I would like to ask Mr. Sun to give a deeper talk. For example, you and your competitors have all invested a lot of money. That day Shen Yang made a joke. In the future, what direction will the industry develop?
Sun Jian: Computer vision industry is not like natural speech processing. In fact, speech recognition is generally considered to be relatively simple. However, computers are diversified. Image recognition is only one of them. It includes a lot of aspects, including sports, emotions, and various aspects. A variety of applications, as well as a variety of applications on the platform, including applications like the new iPhonex phone unlock, very many. So I think that although there are startup companies in the field of computer vision, it is not the same as a homogeneous competition with a single purpose such as trickling or sharing a bicycle. Because of the scenes that it uses, I think Yamashita We will share it. Actually, we have a lot of things that are different.
Shan Shiguang: I strongly agree with Sun Jian. AI in this area is a bit more fragmented than other shared bicycles. Divide refers to the fact that the demand is not a special standard. It does not mean that after I do something, the others can use it immediately. . Even in computer vision, even including subdivision to face recognition, I really think there are so many possibilities that there will be multiple companies. We seem to be a face recognition technology, but if we want to subdivide the scene behind it, there will be more than a dozen different scenarios. Each type of scenario may require you to accumulate different data. It may even change the algorithm accordingly.
As another example, computer vision is more widely spoken. Why do you say there are so many companies now? It's because the demand is really varied. After each demand comes in, everyone needs to rush to get data. A group of people engaged in data, a group of people to engage in algorithms, but also training before deployment, the process is relatively less standardized, not to say that you can do a product, the other immediately can follow.
I have recently given an example. We have a client who made a patrol robot and patrols in the community. Both the owner and the property think that this thing is of little use. But they think that there is something that can be useful. It is to see if the patrol robot can monitor the shit. After the dogs and cats in the community are pulled out, they cannot clean it in time. They are uncomfortable to the owners and they are not happy. This matter is obviously not readily available, many companies can do it, but it takes a long time to collect data, and then to adjust the algorithm, this process is faster, especially smooth and also takes three months to six months to complete this matter This sounds ridiculous. Today is the shit. Tomorrow is the plastic bottle. The day after tomorrow is the cabbage helper or garbage bag. If every one of us spends half a year doing it, we can calculate that there are all things in the world. How long is it for six months? This practice is also currently limited by our level of technological development, making our technology development cycle very long. There is no way for us to have a lot of people. It may be very fragmented, so it may require a lot of people to do such things before at least a tool that is not particularly powerful in the short term comes out.
Just now, Yang also shared that many people may think that our company is a company for face recognition. We call it as the extension, and this “top†also means to expand. One of our directions is to develop an AI development platform for the computer vision field. This platform we call Trainng as A Service is a fool-like algorithm development platform. We hope that customers can take their data and use our platform. , what they want. In this case, this popularity of our AI may be accelerated. This may have completely changed into a different track and is no longer a purely computer-visual technical, service company track.
Dai Wenyuan: Actually, I have been reluctant to talk about funding. Although we have a lot of teams, including us in the company, I'm not willing to talk about this issue purely. This doesn't make sense because we are not doing bicycles. The fourth paradigm is that of a bicycle company. The question we have to discuss every day is how much money we have. We have to prove that we have got more money than our competitors. This is the core competitiveness. But AI does not say that the cash on the books is the most competitive, and that financing is equally important. In my opinion, the AI ​​company is more important than the financing amount is who you take the money, how to get the money. Actually, AI is currently a To B. To C is a very brutal competition. To B finally got faster growth. It is important how you integrate resources. At this time, it is not a question of how much money you take, but who gets the money and how to integrate resources. This is one aspect.
On the other hand, there is more talk about bubbles now. In my opinion, all the money we get from these AI companies is far less than the value we create for the industry each year, not one or two orders of magnitude. We are willing to serve some of the profits our customers create, ranging from a few hundred million to as many as ten billion, which is nothing compared to the amount of money we are currently valuing. So if you look at the whole industry, I don’t think it’s a question of bubbles, but you should also say a lot. Of course, there will always be a company that will exist in any industry. If you look at a company, it will be bubbled or even closed down. However, looking at this industry, I think there is no bubble at all, and even two or three orders of magnitude.
Gu Wei: Actually, I really still agree with what Mr. Dai just said. Because we used to have a report in the past, that is, if the traditional data analysis service, it will bring more than 10 times the return, we used to talk about the traditional subject to data, subject to calculations, subject to the algorithm. Now that we have artificial intelligence, we talk about how many times the order of magnitude can be achieved under the conditions that all these elements are already available. From the perspective of the capital market, I think it is a big trend to be optimistic about this piece. Just now I mentioned that it is because of the existence of this major trend that we are also very optimistic about the development of artificial intelligence in various fields. Our Shield is more focused on the development of the financial sector. We also hope that in the various industries of artificial intelligence, we will have a larger mega company in the future and be able to make better and more contributions.
4. Will people be replaced by artificial intelligence in the end?
Audience question: IBM has a robot called Watson. For example, if you give a hospital a robot to see the patient, will the doctor be replaced?
Sun Jian: Medical care is very complicated and long. It should be replaced in the short term.
Shan Shiguang: I never thought it might be replaced. For now, there are so many specialties in medicine, many departments, and very detailed. This involves the experience of a large number of doctors, and how a large amount of data can be used to form a very good treatment or a procedure for diagnosis and treatment. There may be too much time behind this. It may be more at this level of the text, understanding this level, and involving too few other perspectives.
Dai Wenyuan: My point of view in the medical field is that from a technical point of view, there is absolutely no problem in surpassing doctors. You can follow us on the 14th of last month and the Ruijin Hospital released a diagnostic model for pre-diabetics. We have done an assessment that has two to three times higher diagnostic accuracy than the best experts in the field of diabetes. There is absolutely no problem in this regard.
On the other hand, it is the same problem as the unmanned vehicle. There is no technical problem at all. However, when it comes to promotion, it is a big problem to let a robot go to see a doctor. It's like unmanned vehicles, from a certain point of view, it seems that people have improved their security more than people. But you think, if all the cars in the world are opened by a company, how many people will it kill every day? By the same token, if doctors all over the world are provided by a single company, how many people does this company have to cure each day? This is also a consequence of no way to bear it. So there is no problem in terms of technology, but in terms of promotion, because the entire order may be reconstructed, what is the future medical order? This is because I have no way to comment. It needs to be explored by the industry, including the entire country.
Gu Wei: I think General Dai said I particularly agree. From a technical point of view, it should be no problem. It must be a certainty and has this ability. Including the diagnosis, a company was also introduced two days ago. Their so-called intelligent medical diagnosis system can collect relevant information and data, and some models will help you diagnose. One of the diagnoses is to bring a lot of new doctors. There is no doctor with a lot of experience to help him do the right diagnosis. In addition, even if you are an experienced doctor, he can help you to assist in rapid diagnosis, which is now done. So overall speaking, if you go forward, that is how to better integrate people and machines, including some customer experiences. In general, this trend is correct.
Shan Shiguang: I think it is necessary to add that I personally do not agree with the conclusion that there is no technical problem at all. Now it is a very big problem. Just now, from the perspective of medical treatment, there has been a lot of progress in recent years. There are a lot of problems that haven't been solved yet. Is it really that the current algorithm can really solve it? I don't think it's a matter that is still on board, but it will take time to test it.
Dai Wenyuan: We may not be the same in specific scenarios. There may be some areas. There are still certain technical problems. These areas are in some areas where patients are particularly rare. We will also find that the amount of data is not enough. Of course, this is all. It is a matter of time. In the scenario where the vast majority of patients are sufficiently numerous, technically no problem has been encountered. My point of view, plus one of the previous restrictions is this.
Sun Jian: We can understand why there are so many misunderstandings about artificial intelligence. Everyone has a lot of opinions.
Zhang Xiaodong: I know that the guests on the stage are also full of ideas. I'm sure that the audience under the audience has more problems, but the time is limited, and there are more exciting content to share with everyone. Thank you ladies and gentlemen on stage and thank you Netease. I hope that I will have more time to communicate with you. Thank you!
High purity Tin wire is a special target material, Vacuum coating target technology was used in Automobile glass.
Pure Tin Wire,Tin Wire,Tin Solder Wire,Tin Plated Copper Wire
Shaoxing Tianlong Tin Materials Co.,Ltd. , https://www.tianlongspray.com