AlphaGo's defeat of Li Shishi's achievements made the industry awe-inspiring to artificial intelligence. Everyone is also curious and puzzled about deep learning. Excitement also faces difficulties in research. Situation in Artificial Intelligence and Robotics Summit team AlphaGo behind DeepMind member, Computer Science Department University of Oxford, Oxford-DeepMind Partnership Leader, AAAI, EURAI Fellow Michael Wooldridge Michael Wooldridge, and future research on artificial intelligence Achievements and challenges were shared with the report.
Michael Wooldridge said that at present, there are still many limitations in neural network artificial intelligence, first of all because they cannot understand human social relations. For artificial neural networks, such as AlphaGo, we cannot understand how it thinks. However, artificial intelligence must live up to human identity, and it must have such understandable transparency. but,
However, conscious machines will appear, but before it appears, the world will have a lot of signs of machine consciousness first, and we won't be able to do it overnight.
Which technologies are feasible and which are not feasible?For example, deep learning and machine vision, how can we combine all these technologies to achieve artificial intelligence? What are the chances of success? How do these techniques help us achieve long-term artificial intelligence goals?
Strong AI is very powerful, but weak AI is the current pursuit
This is a very important distinction between weak artificial intelligence and strong artificial intelligence : Strong AI is universal and belongs to what we see in Hollywood blockbusters - for example, the sky roaming 2001 robots, these robots are self-conscious It is autonomous, it is just like a person with various functions.
However, this is a distant dream. It is still a long way to go before this step is taken. At the same time, this is not the direction of our current research on artificial intelligence. At present, most of the research is focused on what we call weak AI.
Weak AI is only its goal is not so high and far, weak AI makes the machine and the computer do something that only the human brain or the animal brain does now, so the weak AI is focused on the specific task - I certainly Knowing that weakness does not mean that it is useless. Weakness is not as easy to do. It is just that targeting is different, so weak AI is mainly aimed at very specific tasks.
What is the computer doing now?
What the computer actually designed meant what to do. Computer or computer, if it is not programmed, in fact, the computer is based on some precise instructions to calculate the machine, it can be executed in accordance with your instructions, the implementation of very fast, it can do tens of millions of data in one second, hundreds of millions of data . However, everything a computer does must be broken down into simple, low-level instructions that are very precise. So if you go beyond this, your computer will be powerless. However, artificial intelligence must be summed up in this directive. So, for now, computers can do those things. What can't they do?
The computer is very easy to do arithmetic, for example to solve some tasks, artificial intelligence can basically do it, it is more difficult down. Computers can do arithmetic and do it quickly and accurately, because it's easy to express arithmetic formulas as such low-order instructions, so arithmetic is very simple, but it's harder, for example solving complex problems - driving, this It has also been tackled recently, but it is basically a problem solving task. Each task is to break this task into simple instructions for the computer to execute.
Does a conscious machine eventually appear?
Further down is a conscious machine. Why is it so difficult? For example, complex reasoning, such as playing games, playing a board game to do complex reasoning, dealing with unclear problems - the computer is executed in accordance with very precise instructions, it is executed quickly but you should be clear about this instruction; In relation to perception, perception is to understand the physical world around us.
But this perception is difficult for computers, and perception is the hardest part for self-driving cars. For example, building a car is relatively simple at present, as long as you know the rules of driving it is easy, but the question is how do you perceive the surrounding environment while driving? The other is judging, judging is that there is no precise rule, and many times to look at intuition, conjecture, this is very difficult for the computer.
We say that strong AI is a self-conscious object. I think that strong AI should not be realized in the short term, and I do not think that events such as AlphaGo will lead to strong AI. In other words, the machine can play chess, it can recognize faces, do a lot of tasks, and it does not mean that it does not mean that it is conscious.
What is consciousness? We also have no way to identify a conscious machine. So this conscious machine will not appear overnight, but it will appear in the future. There is no magician. There are many conscious signs before the appearance of a conscious machine. This kind of consciousness is introduced through this sign. For example, if we come to this meeting today, this will require a kind of consciousness. I think this breakthrough is very important, and I think this has a transformative effect on society. Ultimately, this kind of technological advancement makes people healthier and makes humanity more With the ability to do more efficient activities, government and business are more efficient, and artificial intelligence will bring us greater benefits.
The Difference and Integration of Neural Network Type AI and Symbolism AIToday's artificial intelligence has two methods: one is unfashionable, and the other is fashionable— neural network artificial intelligence and symbolism artificial intelligence.
The artificial intelligence of the future must combine the two types. The nerves I talk about are some neural structures. This kind of artificial intelligence is inspired by the human brain's nervous system. The other is the symbol method. The symbol method does not mean copying. The structure of a brain, on the contrary, we express the reasoning ability of the human brain in terms of conformity.
Artificial intelligence deep learning and the advantages and disadvantages of neural AI
The AI ​​of a neural network actually depends on the inspiration of the human brain's micro-structure. Let's look at the structure of the human brain. Basically, we input an idea into a neural network and enter an implied neural unit. If it obtains a certain configuration, it can choose from it, and each input has different weights during the selection process, and the calculation is performed according to this weight. This input represents the environment around us, and the output reflects the choice of our behavior.
This is a very old idea. It was proposed as early as the 1940s. In the 1950s, there was research on artificial intelligence in this area. However, it disappeared in the 1970s and there was a new breakthrough in the 1980s. This is a very long process. In the past decade, there has been a lot of research in this area. Because there is a true technological breakthrough, this neural network requires three elements:
The first one needs to have a breakthrough in this algorithm.
In 2004, some scholars at the University of Toronto and the world put forward some basic and fundamental new technologies to organize the neural network. But most importantly, they can also build a large amount of training after this neural network. data.
The second is training data.
In the past 10 years of big data, the data is what we have now. For example, we have social media. You took a selfie and wrote a name. It means that the social network has acquired your information. This information helps us objectively. The network is trained to identify other people's faces.
The third point needs the ability to operate to be able to train this neural network.
So this is exactly the development of the other in the past decade, and I think that these three factors have led to a major breakthrough in our recent deep learning and neural networks.
For everyone to see a short video, it is DeepMin - this is the previous program, this program is playing video games. They didn't play well in 100 training sessions. They didn't know what they were doing, and they did some random actions. But after slowly passing through training, it plays better and better. After 400 trainings, it was basically like a human player.
The developers of this game did not predict such behavior before. This is a completely learned behavior. This program learned how to play this game? It found that the best way to play this game is to do just that. The most important thing we see is that this program does not know what it is going to do at the beginning, it does not know how to swim this game, it sees what we see with the human eye, and then it starts to test to experiment and play differently. The way to learn to get more points.
For example, if you want to play a game, you need to have 200 different possibilities. If you say that you have to go two steps, you must look at 40,000 different possibilities. If you take 10 moves, you need to see 1 times. 10 of 23 possibilities. If a computer program is to look at 1 billion steps, it may take 3 billion years to assess all possibilities. So now Intel’s fastest processing capacity will not help you to overcome such problems. , you need to have other skills, how to do it? Monte Carlo search tree.
Using the Monte Carlo search tree, AlphaGo is indeed a major breakthrough for our artificial intelligence, but even though it is a huge success, it does not allow us to implement universal artificial intelligence. What does this mean?
Going back to our conscious and subconscious actions, this is not in Alpha Dog's experience. We don't know how to implement universal AI. It doesn't explain the strategy he uses. Even if it plays this Go, he can't tell you. To explain this trick, we can't extract the strategies it uses in the Alpha Dog process. What can't it do with the Alpha Dog system?
A conversation like this one can be seen on a television show or a movie. He says I want to leave. He answers who he is. You want it to explain what happened. We might say that Ann would break up with him. Then Bob thinks that there is a third person. This is from the person who can draw this conclusion, but how did you come up with this explanation? You have to know some of the theories of human relations. You also have to know this kind of mechanism for the relationship between people. And you need to have relevant background knowledge to get reasoning, but artificial intelligence is very difficult for us to get this way. An explanation cannot be introduced with such a simple neural network. It is indeed our lack of such knowledge of human and human relationships in computers.
This is another sentence. This is two sentences in English. The first sentence is that this committee has rejected this group to march because they are supportive of violence. To this sentence you will ask who they are referring to. We will know this. They mean the group you want to carry out. How do you come up with this explanation, because the committee generally does not support violence, so if you want to come up with this explanation, you need to use knowledge of the Human Organizing Committee. The second one The example is that the committee refused permission to send demonstrations to groups because they were afraid of violence, so the first sentence they refer to the group, and the second they are referring to the committee, so we have to understand these two sentences you must have Knowledge, but also knowledge of human society.
Another example is machine translation. You may see that it is a translation of the original text, but it cannot fully understand the translation of the original text, and it cannot achieve faithfulness. It is also unacceptable.
The other one is Van Gogh's paintings. How do you explain the mood of this picture? Art students can say what kind of mood it represents. If you put in Microsoft's best graphics recognition software, how do they interpret it? They say they don't understand, but it seems that there are two animals swimming in the water.
The inability to understand human social relations is the lack of machines.
Artificial intelligence symbolism make up
To let the machine understand human social relations, I would like to talk about another area of ​​AI research. It is not fashionable today. Today everyone is talking about deep learning neural learning. This is called symbolic AI.
We humans are grammatical reasoning animals. For example, my wife, I think she is happy. Red wine can make her happy. So I bought her red wine. This is a very clear reasoning process. But if we turn it into a symbolic AI, we let the machine reason in the same way, let the machine reason the same conclusion.
There are many obvious benefits to the symbolic artificial intelligence:
The first benefit is its transparency. AlphoaGo is doing well, but it also has no way to explain how well it works. But here in symbolism, we can clearly explain why the machine does this. This transparency is indeed an architectural advantage, and it is the benefit of symbolism AI.
The second is that this level of knowledge is actually very close to our human language . It also uses a very specific reasoning method. If Alphoa Go has no way to explain how it plays chess, it will use the language's ability to express its ideas in sentences.
Of course, why symbolism AI is not popular, it is flawed.
The first is conversion capability. For example, in a complex reality, it is very difficult to express complex and simple symbols. This description is expressed in a simple sentence.
The second question you need to reason about this process is to express this process of reasoning. This is extremely difficult at present.
The next challenge for AI is to tie together different dimensions
I think the next challenge for AI is to tie together different aspects. First of all, we have a very sophisticated program. For example, a program like AlphoaGo has been very successful. It can do a lot of complicated things, but in a sense it is a black box. It has no way to tell you how it won. , so we need a level of knowledge that can be raised above the neural network, summing it up into a language and an understanding.
We can use the language of (using consensusism) to use and communicate, AlphoaGo has no way to explain itself, there is no way to explain why it does this, and it can not self-examination, so I think the ability to explain this is extremely important. For example, DeepMind is mainly engaged in medical research. It is necessary to explain why the patient must have the ability to explain why he wants to eat the drug. Therefore, it is a key point in using it.
Questions and concerns about artificial intelligenceI want to emphasize: The recent success is very narrow and it is only a success in a particular area.
We will not see singularities in the short term.
The so-called singularity is to this point the machine broke through its level. We think about the economics of artificial intelligence: the machine is now beginning to replace people. So if things let the machine do, what should people do? In particular, the increasing level of machine receivers will lead to unemployment and unfair problems.
There is also privacy, such as face recognition, which also has the problem of personal privacy, and society must face this challenge in the next 20 years.
Finally, there is the question of automatic weapons. What will be brought about by the use of artificial intelligence in weapons during the war is not military personnel but machine military personnel. We must think clearly in this regard.
Finally, I would like to talk about if everyone is worried about the machine in charge of the world, I invite everyone to look at a short film. This is a short film from the United States. The Advanced Institute of American Gold has a robot contest. This machine can't even open the door, so if One day when the robot takes over the world, you close the door and the door doesn't have to be locked. It won't come in if you close the door.
Question: In the future when human beings communicate with artificial intelligence, it is impossible to understand the internal thinking logic of the machine. It is as if you can understand the internal thinking logic of the machine when you talk to others.
Michael Wooldridge : I think this is a very good question. This is also a key issue. To be recognized by human society, artificial intelligence must have transparency. It can explain why I do what, why should I do this? So from a lot of challenges, it is difficult for computers to get people's acceptance, but to accept it must be logically transparent. As such a symbolic AI is easily clarified, but no one knows you This kind of behavior is based on the kind of logic. This is a big challenge. We must overcome this challenge. In this case, the artificial intelligence technology can understand the identity.
Question: The first problem is that from the perspective of investment, artificial intelligence has already had a lot of enthusiasm in industrial applications. Where is the next hot spot for investment in the artificial intelligence industry? The second question, because both of them are directors of the joint laboratory of industry and academia, from their point of view, the experience and inspiration of cooperation between industry and academia.
Michael Wooldridge: The next application of AI, especially in industrial applications, I think the next big application should be in the medical field. Why do I think it will be in the medical field? Because in the United Kingdom and the United States, I believe that it is the same in China. Now it is very popular like this bracelet. It can monitor your heart rate, your blood sugar level, and how many steps you take, like an Apple Watch. The same is true of this wearable device, which is constantly monitoring your physical condition. All this information will be given to the AI, so that it can be used to achieve some healthy applications.
I don't know what happened to you. I only went to see a doctor when I was sick. Now this application is for the doctor to be with you at any time and monitor it 24 hours a day. It knows how much you sleep, how much you eat, and knows your blood sugar. At the level, he knows the status of your exercise through the skin, etc., so the smart phone can advise you when you have to exercise, when you eat too much, or drink too much alcohol. In the UK, they said that their biggest application is in health care. They now have records of this, that is, from the British National Health System, they have records of this case, including all British cases, medications, and The AI ​​we talked about this morning can actually be applied to the medical field. This can really bring new and more discoveries to our entire medical industry. So I think the next hurdle should be medical treatment.
The MTP series DC Power Supplies are High Power DC power supplies developed by iDealTek-Electronics based on IGBT semiconductor components, adopting AC/DC power processing topology with full-bridge phase-shift soft-switching technology. This series of High-power DC power supplies have the most output models and the widest output range in the company's DC power supply series, from 3KW ~ 30KW integrated in 19-inch standard rack-mounted high-precision DC power supplies to 45KW ~ 2000KW with casters or floor-standing Cabinet type High Power DC power supplies, output voltage up to 2000VDC, output current up to 6000A.
This series of High-power AC - DC power supplies are equipped with a reliable two-stage conversion mechanism drive logic circuit and a fast control loop optimized by iDealTek-Electronics. Which balances the requirements of low output ripple and fast DC output response speed, making this series of High-power switching power supplies can provide high-precision, low ripple, high-stability and High-power DC output with fast response speed feature of the switching power supply. All MTP series DC power supplies are equipped with a short-term 2 times rated current overload capacity (Except for some high-current output models) to cope with the inductive and mixed load's demand for high-current output at the moment of starting.
The full range of MTP series high-power DC power supply adopts digital control circuit. You can set and control the power output through the buttons on the front panel of the power supply (3KW ~ 30KW) or LCD touch screen (45KW ~ 2000KW). The high-precision 4 1/2-digit LED (3KW ~ 30KW) or LCD (45KW ~ 2000KW) provides intuitive, high-precision output display and measurement functions. At the same time, the full range of MTP high-power DC power supplies are equipped with RS485 interfaces as standard, following the MODBUS-RTU international communication protocol, which can realize remote control and operating status monitoring of the power supplies.
At present, this series of High Power DC power supplies are mainly used in Battery charging testing, DC motor testing, Photovoltaic inverter testing, and various cutting-edge applications that require High-power DC output.
High-power DC Power Supplies, High Power AC - DC Supplies, High-power Switching Power Supplies, High-power AC DC Supplies, High-power Power Supplies
Yangzhou IdealTek Electronics Co., Ltd. , https://www.idealtekpower.com