Artificial life and artificial intelligence

The development of robotics is approaching an important stage: the possibility of creating artificial life and artificial intelligence.

Artificial Intelligence

The dream of mankind is to create a machine equipped with artificial intelligence (AI), able to compete or even surpass the intellect of man. It seems to me that the introduction and development of artificial intelligence (AI) in computer systems is best possible through the creation of neural networks. This does not coincide with the opinion of other computer specialists, who consider expert systems and special systems of rules for a “task” (program) to be potentially more viable.

Indisputable is the fact that “task” operating systems (DOS, Windows, Linux, etc.) and the corresponding software are capable of solving practically all known problems today. Without denying this fact, I note that working with neural networks is the most promising for realizing the dream of creating AI.

More recently, it was predicted that the use of powerful parallel processors in combination with neural networks using the principle of fuzzy logic will simulate the human brain over the next ten years. The forecast was too optimistic, nevertheless, some success in this direction was achieved. The second generation of chips built on the principle of neural networks has already appeared on the market. More recently, two companies (Intel Corp., Santa Clara, CA and Nestor Inc., Providence, RI) have created the Ni1000 neurochip together. Model Ni1000, released in 1993, contains 1024 artificial neurons. This integrated circuit contains three million transistors and is capable of producing 20 billion binary operations per second.

The evolution of "consciousness" in artificial intelligence

The presence of consciousness is a manifestation of internal processes occurring in the brain. The origin of consciousness in humans Homo sapiens was the result of the evolutionary development of the neural structures of the brain as a biological system. A billion years ago, the most advanced form of life on Earth was worms. Let's imagine for a moment this prehistoric worm and ask ourselves the question: could the germ (in the sense of the neural structure) of the intellect produce a rudimentary "consciousness"? If this is so, then this “intelligence” and “consciousness” turn out to be similar to the work of artificial neural networks used in modern supercomputers (see Fig. 2.1).

Fig. 2.1. Graph showing supercomputer capabilities

The worm is undoubtedly a living thing, but can it be aware of "itself"? Or is his nervous system an organized ensemble of neurons that reproduce the same “initial” record that was already incorporated in the neuronal structure of the ancestors, thus realizing nothing more than functional biological automatism?

Is consciousness a life?

Such a question includes several: “Is intelligence the mind?” And “Is consciousness the life?” It seems correct to say that the intellect must reach a certain degree of development, the “critical” mass, and only in this case can we speak of the emergence of consciousness . In any case, artificial neural networks are capable of and eventually reach the level of "consciousness". Whether it will happen in 10 years, or in 1000 years - it does not matter; 1000 years is only a moment in evolutionary history. (I still hope that this will happen in a dozen years, and I will be able to see a full-fledged AI during life.) It is interesting that when artificial neural networks acquire “consciousness” and “self-awareness,” can they be considered living beings?

Artificial life

The creation of artificial life (IL) can be divided into three main areas of research: the creation of “neural” robots with autonomous power, the creation of nanorobots (including the possibility of “reproduction”), the creation of computer programs (software). The most advanced type of artificial life on Earth today are computer programs. Robots capable of self-reproduction have not yet been invented, and it will take a long time to wait for the appearance of nanorobots. For this reason, we will focus now only on computer programs IL.

In such programs, “life” exists exclusively in the form of chains of electrical impulses that are generated by the program in the computer’s memory. Computer scientists created a lot of different IL programs that simulate various biological processes (survival, birth, death, development, movement, feeding, mating, etc.). Some of them are called “cellular automation” (clustering), others are called “genetic” algorithms.

The program of cellular automation (CA) was used to accurately model biological systems and to study the nature of the spread of infectious diseases, such as AIDS in the human population. Similar programs have also been used to study evolutionary processes, the behavior of ant and bee colonies, and many other stochastic systems. To generate random processes, special stochastic algorithms were introduced into the programs. One of the interesting applications was the use of spacecraft programs to optimize the size of neural networks used in the host computer. It is hoped that such programs will help to create and “connect” large neural networks for use in supercomputers.

Genetic algorithms (GA) operate in the spirit of the Darwinian theory of survival of the fittest. Two competing GA programs can “meet” in the computer’s memory and mix their binary codes to produce “offspring”. If the “descendant” turns out to be the same or more viable compared to the “parents,” then, most likely, he will survive. Whether these programs are alive is obviously dependent on the definition of life. What if there are programs capable of self-development and increasing their own "software" level? What happens when similar programs are built into mobile robots? What about robots that have learned how to make their own kind?

Nanorobots - are we living beings?

A nanobot is a microbe-sized robot. IBM was able to achieve some success in creating electronic and mechanical devices (transistors and conductors) with molecular or even atomic dimensions. Such achievements instill confidence in the possibility of creating objects of arbitrarily small sizes, so robots the size of a bacterium are theoretically possible.

Some scientists predict that the next evolutionary step will be the emergence of life based on silicon, which will replace carbon life forms on the planet. What we now call electronic devices and robots will become forms of self-developing and self-propagating silicone life.

A bit of history

The progress of computer technology over the past five and a half decades can be called staggering. Created in 1946, the ENIAC computer was a whole mountain of electronic equipment. With dimensions of 30 m in length, 2.4 m in height and 0.9 m in width, its weight reached 30 tons. ENIAC contained 18 thousand electronic tubes, 70 thousand resistors, 10 thousand capacitors, 6 thousand switches and 1,5 thousand electromagnetic relays. The performance of the machine was 5 thousand additions, 357 multiplications or 38 divisions per second. Today, a similar computer sample of 1946 can fit on a tiny silicon wafer with an area of ​​less than 5 square meters. mm

Physicist Robert Yastrow claimed in The Enchanted Loom magazine (New York, Simon & Shuster, 1981) that “first-generation computers were a billion times' more stupid” and more ineffective than the human mind. Today, this gap has been reduced by more than a thousand times. ”

Science is pursuing an unrelenting pace to create an AI, and as I said, it is possible that artificial intelligence will be created during our lifetime. From the point of creation of AI only a few steps to create a machine "superintellekte." Many scientists will tell you that this is only a dream, trying to keep the sweet illusion of the unconditional and final superiority of human intelligence. Without being comforted by such illusions, I can say that the progress in creating AI is uncompromisingly and with unabated pace becoming a reality.

Perfect than us

Do we, as representatives of a rational race, want to create an intelligence that is superior to our own? If we think about this problem, then in the long run we may need it, if only for the purpose of survival. Think about the prospects of the country that first creates AI with an IQ of about 300. Such an AI machine can be entrusted with the problems of improving the national economy, cleaning up the environment, stopping pollution, developing military strategies in case of conflicts, carrying out medical and scientific research and, of course, creating more perfect AI devices. It is possible that the following theory of the development of the Universe will not be proposed by man (as Albert Einstein did in his time), but by machine AI.

Locked cell

Why is it so important to create superintelligence? Will mankind, in the end, find a solution to this exciting problem? Maybe. The need to create a powerful AI can be illustrated with a story I heard or read. I am only afraid that I will not remember the author’s last name, for which I apologize to him. If I distorted the story a little with my retelling, then I apologize for that too.

There are ten chimpanzees in the cage. The cage door is locked. To guess how to unlock the lock and open the cage door, IQ of about 90 is required. Each chimpanzee sitting in a cage was tested and showed an IQ of about 60. Can ten chimpanzees, by joining forces, find a way to open the cage door? The answer is unequivocal - NO. Intellect does not accumulate. If 10 chimps, acting together, would have a total IQ intelligence of 600, this would be more than enough to unlock the door. In reality, chimpanzees cannot do this.

In real life, we are faced with problems such as environmental pollution, the economy, diseases like cancer and AIDS, the search for longevity, and various areas of scientific research that can be metaphorically represented as a “seat” in a locked cell. From this point of view, the creation of super-powerful AI seems obvious. Such an AI may find the necessary keys to “unlock” such problems, which to date remain in principle unresolved. I do not think that such possibilities of AI remain outside the scope of attention of various states. It is possible that the next “Manhattan Project” undertaken in our country (I hope) will be devoted to the creation of a superIIA.

As a race, we are unlikely to be pleased with the advent of machine intelligence, in comparison with which we will feel like a chimpanzee. Scientific fiction writers have been describing the madness of supercomputers with AI for a long time. Such is the computer HAL in the novel “Space Odyssey, 2001” by A. Clarke, such is the central computer in “Terminator” and “Terminator II”. For all future AI creators who have read this book, I have a warning “Do not forget about the switch!”

Biotechnology

Advances in biotechnology in the near future will allow us to change our genetic basis. Based on this, it will be possible to "modify" our brain to increase its intellectual abilities. However, it is possible that such gene modifications will lead to unpredictable consequences for subsequent generations that could be catastrophic. Creating a super-intelligence based on a machine seems more secure, at least for the time being.

Neural networks - expectations versus reality

The possibilities of neural networks from the very moment of their appearance were, perhaps, unduly advertised. Therefore, it is quite easy not to take into account my considerations about the AI, IL, and neural networks, in fact, as many have been doing for a number of years. Although it is also true that the appearance of "humanlike" intelligence was predicted.

If development proceeds at the same pace as in the last 50 years, then I hope that in half a century AI systems will appear, comparable to the capabilities of the human brain.

What is neural networks?

I described neural networks without precise definition. Now I will give this definition. Neural networks are artificial computer systems (based on hardware and software) that function and "learn" based on models created by analogy with the biological systems of the human brain. Such networks can be created on the basis of hardware / software or be purely hardware. Modeling modeled on biological brain structures has led to the successful solution of some particular problems necessary for the creation of AI, such as machine vision, speech recognition and vocalization. Neural networks can be “trained” to perform pattern recognition. They can be taught to read or check product quality through visual inspection of products. One such example is the Papnet system described in Chapter 1. Other networks can be trained in the recognition of sound commands (speech recognition) and speech synthesis. Networks using statistical methods can predict the behavior and probabilities of events in complex non-linear systems based on past experience. Such systems are able to provide the dynamics of oil prices, ensure the control of electronic devices of the aircraft and predict the weather. Neural systems can also be successfully applied in analyzing the state of the market, evaluating candidates for mortgage loans and life insurance, showing better results than traditionally used expert systems based on standard decision rules.

What is artificial intelligence?

A legitimate question, isn't it? Of course, the development of neural networks will lead first to the emergence of "intelligence", and then - "consciousness". In an attempt to create networks that are intelligent or demonstrate intelligence, what criteria should be used to understand that the goal has already been achieved?

Alan Turing, a British mathematician, proposed an interesting procedure that is generally considered reliable for determining whether a machine has intelligence. The man and the car enter into the conversation by sending messages by teletype. If a machine can communicate in such a way that a person is not able to determine who is on the other end of the teletype line: a person or a car, the machine is defined as “thinking”. This procedure is called a Turing test and is one of the criteria for determining AI.

Although the Turing test is generally accepted, it is not the final procedure for determining AI. There are a number of completely "stupid" language programs that are practically able to undergo this procedure. The most famous of these is the ELIZA program, developed by Joseph Weizenbaum at the Massachusetts Institute of Technology. ELIZA imitates the work of a psychologist, and you can talk with her. For example, if you sent ELIZA a message that you lost your father, she can answer: "Why did you lose your father" or "Tell me more about your father." These answers may make you believe that ELIZA understands your speech. It is not, of course. Answers are cleverly constructed statements based on your posts.

Thus, if we want, we can drop the Turing criterion and take something else. Perhaps the best indication of the existence of intelligence is the presence of consciousness or self-consciousness. A machine that is aware of itself will know for sure that it is reasonable. Another possible criterion, more direct and simple, is the ability to learn from experience. This criterion is used in this book.

Of course, we can discard all logical grounds and argue that intelligence is intrinsic to systems that have a developed sense of humor. As far as I know, apes are the only creatures that can laugh. Perhaps the presence of a sense of humor and emotion will be the best criterion and put an end to his search.

Using neural networks in robots

So, how are neural networks used today in robotics? Yes, we are still far from creating a sufficiently “reasonable” AI, not to mention supplying them with one of our robots. Nevertheless, in many cases, the use of neural network technology allows you to create a system of control functions of robots that exceed the capabilities of standard CPUs and software. Using neural networks in our robots will allow them to perform small “wonders” without using standard computers, processors and programs. In Chapter 6, we will build a fuzzy logic system consisting of two neurons that can track the direction of the light source. Подвижный робот, снабженный такой системой, оказывается в состоянии следовать за источником света в любом направлении. Также в главе 6 мы обсудим технологию BEAM и идеи Марка Тилдена, создавшего транзисторные схемы (нейронные сети), которые обеспечивают движение и иные функции роботам, имеющим «ноги». Большой прогресс достигнут в применении другого нейронного процесса, названного предикативной архитектурой, использующего метод предикативной (условной) реакции на стимул.

Микросети

Небольшие программы нейронных сетей могут быть осуществлены через микроконтроллеры. Более полную информацию о микроконтроллерах» можно найти в главе 6.

Нейронная поведенчески-ориентированная архитектура

The principle of constructing a behavioral-oriented device architecture, developed by Walter Gray, shows that the relatively simple stimulus-reactive neural systems built into the robot demonstrate a highly organized, complex system of behavior. Predictive architecture as a special case of behavioral-oriented architecture was developed by Dr. Rodney Brooks (MIT) and