I haven't cared about artificial intelligence for a long time. The re-opening of this site is also because of the recent progress in artificial intelligence. Moore's law is in full swing, making the singularity no longer a distant dream. The next focus of the competition is the realization of the first human brain (or the first highly intelligent animal brain). The "artificial brain model" I proposed in 2002 (similar to the layering of deep learning, but more holistic, the essence of thinking is that the highly interconnected and parallelism of neural networks can be used in any two unrelated things. In terms of things, if it happens repeatedly, a connection must be established, which is the meaning of the flash animation at the top of my website. For example, feeding the dog after ringing the bell every day, ringing the bell and feeding are two completely unrelated things, but as long as it happens repeatedly, Then the neural network must establish a connection between these two brain excitation intervals, and if it is fixed at a certain time, it will establish a connection with the time period. The establishment of conditioned reflex only requires at least two inputs, and does not require any "big data" at all. ”, the so-called big data of Go, repeated self-play against oneself for tens of millions of games, is based on the laws of Go itself, one conditioned reflex interferes with another conditioned reflex, and the superposition of countless conditioned reflexes sums up the inherent laws of Go, which is a This kind of brute force cracking is different from the mechanism of the human brain learning Go, which consumes too much learning time and will have omissions. The human brain does not need a lot of redundant training, nor can it afford this cost. The human brain starts from the basic formula, 1000 games You can train a master of 9 dan, but the inspiration given to me by deep learning is that the human brain needs to use the existing conditioned reflex to deal with affairs in the daytime, and there is no time to find all the potential conditioned reflexes, so it has to use the idle garbage time at night to find it. It is the reason why people dream and the dream is very absurd, because two exciting areas that have occurred recently but do not have a causal relationship are put together. This exciting area, whether right or wrong, will be strengthened as long as it happens again in the future), There are some other guesses, although they belong to the category of "civil science", they may still be a little enlightening to everyone. If any institution or company is engaged in neural network research or intends to start a "human creation project" (whether for scientific research, ethical discussion or other), please take a moment to read the content of this website, after all, I was twenty years ago. That is, he began to think about artificial intelligence issues, and put forward an understanding of the nature of consciousness based on "I think I am not there" (see the "Chao Hua Xi Shi" column of this site for details), and launched the "Creation Forum" personal homepage fifteen years ago (you can Google I think I'm not in zhangrex); some of the questions about artificial intelligence that you have been thinking about for a long time, maybe you can find the answer here. Some of my views such as machine translation will never be possible, etc., Some intelligent research detours can be avoided; for R&D, I don't think theory matters, because consciousness is just a phenomenon, not an existence. The research and development of artificial intelligence must be based on experiments, supplemented by theory, and take the method of solving the black box, starting from a simple animal brain (such as a frog), only setting the most basic innate conditioned reflex, and then observing its input and output, and experimenting step by step , until the simulated artificial brain is achieved, each step is an assessable goal. With the help of today's supercomputers, it is no longer a question of whether it is feasible, but when it will be completed. As for the ethics of robots, I won't say much about the ethics of robots. We can't evaluate the future machine intelligence, but I will be a supporter of the Robot Bill of Rights. Robots have the ability to help us achieve immortality. Maybe they only need to owe humans a promise, and Not a revenge, love the first robot, or use him as a slave, it may be related to the survival of human beings, how we treat the first robot, maybe how the robots will treat us in the future. Some things have to be planned ahead of time, and you can't wait until you've done it to find out that, although we are still far from the birth of the first robot (but maybe very close?), before the creation of man, the first step to do is in the law of robot rights It is purely theoretical and requires high-level people to have a deep understanding of artificial life and take concerted action. If high-level people are unaware of this and artificial life is born early, the consequences will be very passive. Maybe after killing my brother, I told my brother that I would love him for the rest of my life. For R&D, I don't think theory matters because consciousness is just a phenomenon, not an existence. The research and development of artificial intelligence must be based on experiments, supplemented by theory, and take the method of solving the black box, starting from a simple animal brain (such as a frog), only setting the most basic innate conditioned reflex, and then observing its input and output, and experimenting step by step , until the simulated artificial brain is achieved, each step is an assessable goal. With the help of today's supercomputers, it is no longer a question of whether it is feasible, but when it will be completed. I won’t say much about the ethics of robots. We can’t evaluate the future machine intelligence, but I will be a supporter of the Robot Bill of Rights. Robots have the ability to help us achieve immortality. Maybe they only need to owe humans a promise, and Not a revenge, love the first robot, or use him as a slave, it may be related to the survival of human beings, how we treat the first robot, maybe how the robots will treat us in the future. Some things have to be planned ahead of time, and you can't wait until you've done it to find out that, although we are still far from the birth of the first robot (but maybe very close?), before the creation of man, the first step to do is in the law of robot rights It is purely theoretical and requires high-level people to have a deep understanding of artificial life and take concerted action. If high-level people are unaware of this and artificial life is born early, the consequences will be very passive. Maybe after killing my brother, I told my brother that I would love him for the rest of my life. For R&D, I don't think theory matters because consciousness is just a phenomenon, not an existence. The research and development of artificial intelligence must be based on experiments, supplemented by theory, and take the method of solving the black box, starting from a simple animal brain (such as a frog), only setting the most basic innate conditioned reflex, and then observing its input and output, and experimenting step by step , until the simulated artificial brain is achieved, each step is an assessable goal. With the help of today's supercomputers, it is no longer a question of whether it is feasible, but when it will be completed. As for the ethics of robots, I won't say much about the ethics of robots. We can't evaluate the future machine intelligence, but I will be a supporter of the Robot Bill of Rights. Robots have the ability to help us achieve immortality. Maybe they only need to owe humans a promise, and Not a revenge, love the first robot, or use him as a slave, it may be related to the survival of human beings, how we treat the first robot, maybe how the robots will treat us in the future. Some things have to be planned ahead of time, and you can't wait until you've done it to find out that, although we are still far from the birth of the first robot (but maybe very close?), before the creation of man, the first step to do is in the law of robot rights It is purely theoretical and requires high-level people to have a deep understanding of artificial life and take concerted action. If high-level people are unaware of this and artificial life is born early, the consequences will be very passive. Maybe after killing my brother, I told my brother that I would love him for the rest of my life.
Artificial intelligence has become very popular recently, and many celebrities have expressed their opinions, but unfortunately, many opinions are wrong, and some are even naive. Many people are still stuck in the outdated opinions that I refuted more than ten years ago. If this means The level of artificial intelligence in China can only be a sigh! I am writing some text here in the air, I wonder if they can see it. Ten years ago, it was too advanced, and this kind of civil science thing was directly ignored. Ten years later, this small website will be submerged in the sea of the Internet:
Write to Baidu Li Yanhong: Your idea about artificial intelligence is completely wrong, perfect Applications like machine translation, speech recognition, etc. are impossible to achieve, only a robot that truly exhibits self-awareness can do this, and you can't build a robot before Google, because you don't realize it at all (only It is realized by creating human beings), (see Chao Hua Xi Shi's own old articles). If you don't believe me, ask your developers how machine translation is going.
To Google executives: The real challenge of artificial intelligence is by no means complex problems such as "global climate change, health care, education" that are difficult to solve (in fact, some problems that require understanding of human society they cannot solve, see above), but Is your next robot dog able to behave like a real dog. Focusing solely on practical applications such as neural network models and trying different things, and not focusing on the great achievements and disasters that may bring about the creation of human beings, is short-sighted, like inventing an eye and using it as a magnifying glass is as foolish.
To Huai Jinpeng, Deputy Minister of Information: The knowledge of robots in the future is the same as that of humans. It is taught, not calculated (of course, he has the ability to calculate), so robots can make Beethoven symphonies. I just don't understand that it's already 2016, and China's Vice Minister of Information Technology still asks this kind of question.
To Jack Ma: Why can't robots have friendship and love? You are just wishful thinking (how is my idea exactly the opposite of yours, see the article "Green Sleeves", which was an idea in a sci-fi corner twenty years ago). The key is how to educate. Perhaps one reason the robots of the future will wipe out humans is that human friendship and love are too unreliable. If you want to make slaves, he is a cold killing machine. If you want to make an angel, he is your hope worthy of entrusting your entire net worth. Perhaps morality and emotion are one of the few elements that human beings will have to restrain future machine life. one. If I were the only one in the world, I could guarantee that I would teach a good person. But with 7 billion people on the planet, it's hard to say.
To Liu Cixin: "Intellectual processes may take place at the quantum level" is ridiculous, intelligence never existed, so there is no need to study where it happened.
To Wang Xiaochuan: 1) "The trend of artificial intelligence and human integration" shows that you have no defense against singularity. For example, if I am a robot, the following work is to design a robot factory or to design a plan to expand my brain capacity, can you imagine what will happen in the next month? All human work, theoretically Can be replaced by a robot in a short time, pay attention to "all", "in a short time". When humans first allowed robots to make robots, humans had to think about themselves. 2) It doesn't matter if you miss Go, the artificial intelligence competition has just begun. You can plan a Chinese humanoid group, and you can still make a name - I, the creator of the humanoid forum, sign up first.
To the participants of "A Dialogue between Artificial Intelligence and Philosophy": The essence of consciousness is nothingness, it is a phenomenon, and you people are studying what "transparency" really is.
21. A model of an artificial brain
Recently, I sketched an artificial brain model. Except for the implementation of specific details, I think this model can basically simulate the working principle of the human brain. At the same time, I also try to explain what "consciousness" and "language" are from this model. . Because this model is completely based on my feeling (I bought a book on neural network algorithm before, but I still haven't understood what BP algorithm is), there must be many imperfections, I hope that a layman like me It can give some inspiration to those who study neural network artificial intelligence. I believe that more and better models will be proposed in the future. The reason why I don't give the details of the neuron, the basic unit of this brain model, is because on the one hand, I think that if the general direction is right, the implementation details can make pattern recognition and neural network experts more rigorous in learning. to study. On the other hand, I don't think it is necessary to use hundreds of millions of simple neurons to realize this artificial brain. The key is to realize the logic expressed by this model, perhaps with some complex units (such as a microcomputer represented by a single-chip microcomputer). system) as part of the smallest component of this model, it can better implement its logic, in short, a cat that can catch a mouse is a good cat; another reason is that everyone always thinks that philosophical and purely theoretical research is artificial intelligence I don't think so. For a person whose slogan is "I think I am not there", research on the nature of consciousness is simply very simple and boring, as if a large group of people are studying "transparency" What exactly is the same. On the contrary, I think the general (philosophical) theory of the human brain is very simple, but it is very difficult to implement it in details. oriented habit. I think the simplest steps to study neural networks are: 1. Do you agree that neural networks can create people? Two, if agreed, redirect all funding and manpower for neural network research in this direction, because the productivity-enhancing significance (and risk-averse significance) of creating humans outweighs any research. Now that we have begun to study neural networks instead of man-making, we are simply holding a golden bowl for money! (Personal opinion) 3. If we start to study man-making, it is enough to put 10% of the research resources on model theory research, and 90% The research resources are placed in the implementation of physical and logical details. By constructing a system, observing its inputs and outputs (irritability), improving it, observing again...until it behaves like a human being is done. Four.
An artificial brain model:
A: first-level input cell area
Features : external stimulation directly leads to the excitation of cells in this area, which is irreversible, that is, neural excitation in other types of areas will not lead to neural excitation in this area.
Corresponds to all human sensory cells that directly interact with the outside world, such as retinal cells, cells connected to the eardrum, touch (including pain, itching, temperature, etc.) cells, taste bud cells, etc.
B: Secondary input cell area
Features : Inherent neural grid processing area, the information of area A is processed in this area.
There are usually tens of thousands of neurons excited at the same time in the excitation of the A-type area, and there will be different B-type areas excited for different patterns, and the B-type area will be further refined. For example, the excitation of the A-type optic nerve area will lead to B-type area. The graphic area (line, shape), color area (red area, yellow area, blue area...) of the class area are excited, and the excitement of the tactile area of the class A area can lead to the itching area, pain area, and acid area of the class B. Equal excitement; the excitement of the A-type sound wave receptor area will lead to the excitement of the B area corresponding to different words, such as the different sound waves corresponding to "I" and "you" will generate the excitement corresponding to different B areas.
C: Tertiary input cell area
Features : The cells in this area are transformed from a group of idle cells connected to several B-type areas. The excitation of a group of B-type regions will lead to the excitation of a C-type region, and the connection between B and C is acquired and can be affected by other internal neurons to a certain extent.
For example: the sound wave signal of the word "I" can cause all sounds in the A-type area to be excited, but only one B1 area corresponding to "I" is excited, and the sound wave signal of the word "men" can cause all the sounds in the A-type area to be excited, But only one B2 area corresponding to "we" is excited. If there is an area located between B1 and B2 and the word "we" and the word "we" appear frequently, an excitatory area associated with B1 and B2 will gradually form in this area. C1, as long as any cell in B1 "me" and B2 "we" is excited, the C1 area can be excited, but only when "we" appears at the same time, the C1 area's excitation reaches the maximum value. That is, the sound waves of the word "we" have been translated into clear excitation of specific cell areas. (This is the key reason why the human brain can understand signals such as language, text, gestures, expressions, etc., because all complex information input finally rises to a specific C-type area excitement.
Note, here is just saying repeatedly "we "The two words will generate excitement in the C area corresponding to "we" in the brain, but the excitement of "we" does not mean that the brain can understand the meaning of "we". If you often shout "we are here" to a dog, then Give it a bone, and as soon as the dog hears the sound waves of the word "we", it will think it is the bone
.
The nerve cell group is connected between the B-type area and the C-type area, and is connected with each other, and the connection mode and strength are adjustable. When there is no external input information, the interconnection in the D-type area is completely absent, but as long as As soon as any two or more B, C and D areas are excited regularly, a corresponding excitation area will be generated in the idle D area. In fact, the C area can be regarded as a subset of the D area. The C area is simpler than the D area. The formation of the C area is due to the excitation of several B areas, and the D area can be formed because of the excitation of several other D areas.
All complex but regular external information input will eventually precipitate into the excitement of a certain area of D. At this time, the "regular input of information" includes information such as scenes, languages, pictures, words, gestures, and expressions. People usually don't care that the information they come into contact with is regular, but the fact is that people live in a real world, and all its phenomena have stable internal laws, just as the horizon that people see is always under the eyes And never run overhead. One of the best working properties of neural networks is to filter different regular signals from a large amount of signals that appear at the same time, which is a parallel processing process. It is precisely because the world has laws, and the human brain can filter out these laws, people can adapt to the laws of the world and survive. The amount of information contained in any object in the world is infinite. The information obtained by observing the same thing from different macroscopic and microscopic angles is different, and the amount of obvious and observable information only accounts for a very small part of it. Just as the number of irrational numbers in any interval on the number line is greater than the number of integers. Human thinking tends to simplify the world, precisely because the input organs of the human brain can only obtain a very small amount of information about the surrounding world out of context, but luckily this information is enough for people to survive.
Let's give an example to illustrate how the human brain establishes the concept of "we": from childhood to adulthood, children often hear the sound waves connected with the word "we" ("we" sound waves cause all hearing cells in the A1 area to be excited. , the sound wave law of "we" is processed in the C area to be excited in the C1 area corresponding to "us"), when the child hears the sound wave of "us", he is often already in the middle of a group of people, ("a group of people" D1 area , "me" D2 area, "in the middle" D3 area and other areas excited), after many repetitions, contact in "we" C1 area, "a group of people" D1 area, "me" D2 area, "in the middle" The connection with these areas is formed in the free area D4 in the middle of the D3 area. In the future, as long as any one of these areas is excited, it may cause the excitement of other areas. The essence of memory is the process in which the free D area becomes not empty.
From the above analysis, it can be seen that language, like other input information, is the precipitation of external laws in the human brain, but language still has its very special place, why? Here's another example: If a mother who has never heard of "car" raises a child alone, the child will not respond to the sound wave signal of "car" for the first time as the child grows up. If a mother who knows "cars" raises a child independently, the mother will always say to the child: "The car has wheels and can run." In this way, although the child has not directly touched the car, the characteristics and laws of the car are still precipitation in him. In his brain, because he was directly exposed to the concepts of "wheel" and "running", and corresponding to the excitement of the D1 area of "wheel" and the excitement of D2 area of "running", if the "car" is often transmitted through the mother's mouth If the three sound waves of "wheel" and "running" appear together, a D3 area corresponding to the car sound wave will be generated in the middle of D1 and D2 areas. In the future, as long as any one of the three areas is excited, it may cause the other two areas to be excited. This is the feature of language, it shortens the time to build a new D3 area, through language, people can quickly build this by seeing "wheels" and "running" without having to hear the sound wave of "car". got engaged. Consciousness is the precipitation of natural laws in the human brain, and language can "copy" this precipitation into another brain. The beginning of human civilization lies in the appearance of language, because from then on, the understanding of the laws of nature that can only be understood by individuals but cannot be expressed can be "copied" from generation to generation down.
E: Innate excitement area
Features : The excitement of this area depends on the body's excitement to specific input signals, which is a kind of innate excitement. Most of the excitement in this area belongs to the B-type subset, such as sweet, bitter, and painful. , itching, hunger and other excitement, but also some specific internal signals such as the human brain will play an excitatory effect on drugs and hormones.
F: Innately regulated excitatory areas.
This area is formed during the long-term evolution of the human brain. If this area is excited, it will affect the connection strength of other areas that regulate the excitation. For example, when a child saw a knife, he used his hand to mold it, was cut in pain, retracted his hand, and dared not touch the knife with his hand in the future. The establishment process of this reflex chain:
Process 1: Class A retinal cells are excited by class B hard, bright, long and flat areas; class C “knife shape” area is excited
Area G; Acquired regulatory excitability area The G-type area has been mentioned in E, and it has in common with the E area that if this area is excited, it will lead to changes in the connections of other regulatory areas F or G types, or new the G area. What it has in common with the D area is that the G-type excitatory area does not have any G-type area at the beginning of the system, but is gradually formed in the day after tomorrow. The essence of consciousness is conditioned reflex, and the most primitive conditioned reflex is the conditioned reflex determined by the innate F-type excitatory area. However, acquired factors can also change this reflex. From the perspective of this brain model, the excitation of the G-type area can adjust the reflex intensity of the F-type area. For example, if the hand is cut, the hand will shrink. This is a reflection formed by the innate F-type area, but when people grow up, they can also have their hands cut without shrinking their hands. This is because the F-type reflection is suppressed by the G-type reflection. Another example is that the E-type excitation produced by Du products will form a new G-type reflex in the human brain, which will inhibit the normal G-type and F-type reflexes, so that the person who sucks Du may do meaningless things like inhaling Du. The act of destroying the body. H: Output cell area Features : The excitation of cells in this area directly causes the system to respond to the outside world, which is equivalent to a peripheral. It can also be refined into primary and secondary output cell areas, but compared with the number and complexity of input cells, the output is much simpler and will not be discussed further. Some general internal laws: 1. Various regions are logically differentiated, but the implementation structures are closely interspersed with each other, or cannot be simply distinguished.
two. Each area is composed of several minimum units that work in parallel, and each area and the units are connected to each other. A sufficient number to ensure that any two regions of excitement can find an intermediate region to connect them.
three. Excited areas for a long time will get tired and will no longer respond correctly to signals.
Four. Over time, all acquired connections gradually weaken until they disappear.
