Chapter 1. Signal Handler
In everyday life, we don’t think about why we like steak or orange. He is fried, juicy and meaty, and we are hungry. Thanks to Darwin's theory, scientists, fortunately, can already explain to us why we like steak. Because in the process of natural selection, everyone who simply didn’t like these steaks died. Those who fried meat on fire survived, and managed to sacrifice their huge jaws in favor of a larger brain. The survivors brain is finely tuned to the fact that grilled steak is good. In the brain for eating such a dish, a number of hormones are released that activate the pleasure center. These hormones - dopamine and serotonin, and several more, form a cocktail of a certain concentration, to which the so-called signal processor or interpreter reacts in our brain. Each signal is like music, consisting of a series of notes. Steak - a kind of sum of signals of the nervous system, including visual, olfactory, taste. These three systems give a certain unique amount, which our processor interprets as “good” (this was developed as a result of natural selection !!!), and then sends a “good” signal to our consciousness, so we feel satisfaction. But what exactly do we feel? It is unlikely that someone will be able to describe in words the feeling of happiness. But if sensations cannot be described in words, at least they can be described in mathematical numbers.
So, the conclusions - we are not able to decide for the handler whether some thing is bad or good. In the human body, the role of this processor is played by a small organ called the hypothalamus. The size of this organ is about the nail, and it is connected to the entire nervous system, absolutely to the whole brain. It is this body that decides for us that in the desert we want a glass of water in the middle of the table. Its fine-tuning itself occurs in the process of evolution, and it is he who decides, not us. The simplest example is that a generation was born in parents in which one handler interprets an orange as tasty, and another handler interprets an orange as not tasty. Here we need to make a note from biology that the next generation may have not only signs of father and mother, but also random signs due to mutations. In the future, the accumulation of these random traits gives advantages to some descendants through many generations, and other branches that do not have such advantages either die in the struggle or simply die out without food. Or they coexist in different territories far from each other. Well, about oranges. One of the children began to live longer for 20 years due to vitamins from an orange, and spread its genes more strongly. After 10,000 years, the entire population of people love oranges, because everyone who did not love them became extinct long ago. And what is love for an orange? This is the response of the neural network to the melody of the signals of the nervous system. This is the response of the signal handler, the conversion of the sum of the signals by the handler function to “good” or “bad”. It can be emulated in any modern programming language. Conventional neural networks can identify the orange in the picture. But what if this neural network determined by the picture, is it bad or good? You show her an orange, she answers well, you show a snake - she answers "badly". You can train the network yourself by accidentally changing the coefficients and explaining that an orange is good, but a snake is bad, but you can let it develop itself, you just need an environment.
Chapter 2. Neural network environment
What does a neural network need for self-development? We need conditions that will force her to do this, including motivation and opportunities for self-development. Need natural selection. Let's create the simplest human model in an empty virtual space. To each such model we will tie our own neural network. To begin with, we will teach these bots to walk. You need to emulate about 50 muscles in the body, for starters. By effectively controlling the muscles, the neural network will be able to move through space. For effective movement to reward, giving the pluses of the network. The same pluses that the signal handler gives. This is already emulated today, you can find many Youtube videos where various funny robots learn to take their first steps. But not so simple. Our bot will move in space the more efficiently, the more parameters at the entrance to the neural network we have. First of all, of course, machine vision. In the second place is machine touch. We will divide the surface of the 3D model into small pieces, something like pixel skin, and each piece will have a parameter, whether it touches something or not. So, for example, the feet will conditionally feel the ground, which will give a smoother movement in the final result after training. You also need to add to the input parameters from the virtual gyroscope in the virtual head. 3 parameters of displacement along the XYZ axes, which determine the inclination of the head or the whole body, as well as the acceleration force and its possible vector, for more precise coordination in space. Together, this system will resemble a kind of juniper, which is present in the organisms of all vertebrates.
What's next? Next you need to teach them to breed. If we started emulating people, we need to emulate them completely. For example, we created a virtual man and woman. You can program that the closer their genitals are, the more pluses of satisfaction fall. But as soon as our guys finished all their affairs, the pluses stopped falling for a certain time. Let them not only have sex, we have more important things to do.
According to the same scheme, you need to teach them how to eat. Our bot is already a little able to control its body, try to emulate hand movements to it. Dexterous hands are not such a difficult thing for today's neural networks. Create a virtual orange next to it. And we will mix the weights of the neural network until its signal processor decides that an orange is good. And that eating it through the mouth is good. After that, our virtual creature will eat virtual oranges with its mouth, all that it finds in its virtual empty space.
In this case, the human model will have virtual vision, and get the picture through the eyes, and preferably in stereo for 2 eyes. With a virtual hearing in a certain radius, for example 40 dB within 10 meters and the closer the signal, the louder it is, the farther, the quieter it is. Still need a sense of smell, touch, taste. All this is emulated easily. In simple emulation, the smell can consist of 50 different shades that make up the composition, to which the signal processor responds well or poorly. Each object in the simulation also has such a parameter as smell. Suppose the rotten meat in the simulation has some random coefficients. For a neural network it’s just a number, it doesn’t really care about how it smells, because in fact there is no smell there, there is only a set of factors. Someone likes it, someone doesn't. Those who like this smell eat meat and die faster. So in future generations it will be fixed that the smell of rotten meat is bad. Once again, there are no smells in objective reality, a smell is a purely subjective thing, it is not a material thing, a smell is the ratio of the brain or neural network to a set of molecules or coefficients in the virtual world.
With taste, everything is exactly the same, with touch it is already more difficult, but in itself, touch is not so important in the development of an organism as vision and hearing. Touch is either a method of interacting with something, or it is a method for responding to something. For example, physical pain from a stroke. In this regard, touch is more important. For our virtual bots, you can emulate that with strong blows he will not be able to use any muscles and movements will have to be retrained slightly. Also, when receiving damage, in the form of strikes or something else more dangerous, the signal processor perceived these situations as “bad”, trying not to repeat in future generations. Those who like blows, pain is perceived as a positive emotion. These creatures would quickly fail all muscles and other functions of his virtual organism, which would ultimately limit the possibility of reproduction.
Chapter 3. Simulation of the body
In order to emulate a living creature, it is necessary to deprive him of the regime of God. Make it very fragile, add external hazards. Imagine a 3d model of a person in a simple editor.
Add to it the miscalculation of the internal skeleton that the bone may break. We add the simplest calculation of the amount of blood in the body, its movement, but without much accuracy, just like a series of parameters. For example, there will be 4 liters of blood in the model, when you take damage there is a chance of bleeding. With large blood loss, death, with small blood loss are simply negative signals from the signal processor. You can add to the simulation the ability to cut off limbs, you can not do this. Stupid bots from this will not. Be sure to add the ability to emit and hear sounds of different frequencies and different durations. Let these sounds affect your surroundings and those around you. Excite in others some reactions. If the creatures use sounds, it is more likely to develop a language we understand. After all, what is language for us? What happened before a person used the language? It is logical to assume that at the dawn of mankind we used different sounds, like animals, the easiest way to look at monkeys is to use shouts. Initially, there was a set of different sounds in random order. Some sounds frightened others, some attracted, so in the process of natural selection a language arose, in a completely natural way of development in the process of interaction of neural networks in our heads. Let the neural networks interact with each other, while increasing their chances of survival, and very soon they will learn to communicate in their own language. Initially, the set limits for the frequencies available for communication can make a language look like a human language. For example, dolphins have a semblance of language, but it is located at ultrasonic frequencies - most of their sounds are not even audible to humans. Language is simply a collection of random sounds generated by evolution.
By the way, breathing emulation can take place if it makes any sense. In the same way, give positive signals for each breath. And those who did not breathe died.
Finally, our body is ready to go somewhere and survive, it is time to include natural selection.
Chapter 4. Reason for Consciousness
Imagine a person who has been deprived of signals entering the brain since the womb. From the optic nerve, from the auditory, tactile, gustatory, temperature and olfactory, no signals would ever come to the brain. What would be in the brain of this creature? Vryatli it would even dream, because in dreams we see things that are usually known to us to the extent that we can at least understand them. And if you have not seen anything? And I did not hear anyone, I did not feel another person, the brain seemed to be forever asleep. Consciousness, as we imagine it, would not even have arisen there. How do we want to receive it, without all these signals? The easiest way to get something that has an analogue of our will is to put the neural network in the simulation, give it a body, and thus the reaction of the brain or neural network to the surrounding reality by means of signals of the nervous system is consciousness. Consciousness is the response of the brain to the environment. No environment - no consciousness. It is not necessary to answer difficult problems of consciousness, such as why red is red, why there is subjective experience, why there is I. All these questions disappear when we consider consciousness from the outside as a reaction of the brain or neural network to a set of incoming signals from external factors. Why don't we hear ultrasound? Because the ear is physically unable to respond to such high sound frequencies. These frequencies simply do not cause vibrations in the ear. Moreover, if the ear was physically arranged differently, the brain, perhaps not immediately, but over time, would learn to successfully interpret ultrasound. We should not be interested in what exactly a person hears at this moment in his head. After all, even if we ask him to describe, he will not be able to do it. We just need to understand that it successfully interacts and responds to this signal.
Once again, the signal entering the brain is interpreted by him. Color, sound, anything else. The brain is unable to describe how this is interpreted, i.e. it is impossible to describe how red looks. So it makes no sense to try to emulate consciousness as such. You only need to program the brain's responses to signals.
Chapter 5. Simulation of the environment
Simulation of the environment will give us several possibilities at once. First, neural networks will receive a constant stream of signals to all possible programmed receptors. Even if we take our reality and emulate only 10% of it qualitatively — no atoms, Newton’s simple physics, everything consists of small particles and is based on the programmer’s honest word. In this world, there may not be chemistry, quantum physics. But there should be opportunities to tear fruits from trees, break a coconut with a stick. It is necessary for the creature to freeze and come up with a shelter. It is necessary to emulate the stone age. Let it be warm in the cave, even after many virtual years of emulation, this creature will one day take random actions that launch the bonfire script.
In a quantitative sense, we don’t need to emulate so much, let's start with a few square kilometers, let the creatures live in a world where, after a dozen kilometers, you return to your original place. It seems to be a finite world limited in area. Or just with a perimeter fence J.
The easiest way is to emulate animals and their behavior. But if we want real human behavior, we need something more. At a minimum, we need to describe for ourselves what makes us different from monkeys. It is on these monkeys that the behavior of neural networks will initially be similar. But the deeper the simulation is worked out, the more opportunities its inhabitants have, the smarter they can become. Imagine that at some point these creatures took possession of sticks, learned to slaughter animals and use them to light a fire from stick friction on a stick. Everything is implemented using existing scripts and neural networks. But then we will run into the limit of the development of this tribe. There, additional computer capacities will already come to our aid, the stocks of which are increasing every year. You can emulate all large territories, in addition, you can add real chemistry to the virtual world. At first simplified, then more and more similar to the one that we know. In addition, chemistry has nothing to do directly with the emulation of atoms. Chemistry can exist without simulating real atoms. A small set of chemical reactions for the created world will be sufficient.
Chapter 6. Development
Life is inseparable from death. Therefore, for the development of artificial intelligence, it is necessary to emulate reproduction, birth and death. With each generation, descendants will adopt the traits of their parents and a small percentage of random parameters that will be fixed in generations. There are a lot of parameters to set, but not as many as are contained in our DNA. Even 1000 fractions of what is in DNA is enough for a start. For example, height, weight, strength - all this is tied to the body emulated by us. Most importantly, these parameters do not have to contain all the reactions of the signal processor of previous generations. It is enough to imagine the signal handler as a function that varies with generations and takes some values (good, bad, like that ...) when applying this function to incoming signals. Thus, it is not necessary to remember the answers to all questions (is an orange good?, Rotten meat good?), It is enough to use our signal handler function and measure the result. We spend a million years. We look at the result. Are our primitive people still digging sand for roots? So emulation limits their capabilities only on this. Fortunately, our world allows us much more, and therefore we have achieved more. By gradually increasing the possibilities of simulation, we increase the potential possibilities of its inhabitants. All this can be improved indefinitely. Or you can at some point let such a creature control an avatar in our real world, if it is ready for this. But this does not replace evolution in the simulation, which is many orders of magnitude faster. Yes, and there will not be much sense in this, because while the simulation is not as accurate as our world, then the creatures will most likely be dumber than us. Only in conditions like ours, can a mind like ours be born and not otherwise. So the beginning is set, it's time to start the simulation.