About the problems of artificial intelligence perception of reality
The problem of creating artificial intelligence is already "a hundred years at lunch." And everything is limited to creating a “Turing machine”, which in the end is just a simplified version of the “Chinese room”. But we need not a talker-interlocutor, but an autonomous robot. And here it’s still less rosy. Why? Read on ...
Currently, the problem of artificial intelligence is usually reduced to modeling neural connections by program methods or to algorithmizing all possible situations and answers to them. These are attempts, at one of the levels, to copy the natural brain. In the first case, at the elemental (neurons and connections between them) level. In the second case, on the instinctive (stimulus-reaction; “if - then” construct). Both for the first and the second path, large computational powers are required, which, according to some, can only be given by quantum computers of the future.
It seems to me that the problem is not only in power. The “root of all evils” of this approach is that the developed artificial intelligence systems are “real-time” systems. That is, the stimulus - data processing - is the answer. It seems the animal’s brain works the same way, but it’s not quite so. The brain of a sufficiently developed animal (we will not take worms and lancelets) does not work in "real time", but in the "extrapolating" mode. As a reaction, a certain pre-calculated answer is issued, not an instant reaction.
This is due to both the relatively small processing power of the brain and the large amount of sensory information coming into it. For example, take the skin of a mammal. From each hair comes a nerve, that is, a source of information, and how many of those hairs (in humans from 200,000 to 1,000,000). And everyone feels not only the fact of touch, but also the power of this touch.
In addition, from the moment of exposure to the sensory organs to the response of the muscles, a lot of time passes (tenths of a second) during which the body decides to live (to catch prey) or die (to miss prey). Our brain (as well as the brain of other higher animals) is forced to live in the past.
To circumvent these problems, evolution has found the following solutions:
1. Signals from the senses are not immediately transmitted to consciousness. The brain is forced to filter them at the subconscious level, thus bringing to consciousness only what is “important”.
2. Pattern recognition is carried out at the "screen" level. That is, slightly similar - it means it. To further accelerate the response of the body, information is not seen in a detailed picture, but discrete, simplified images. Moreover, the degree of simplification increases in stressful situations. In humans, this is expressed in the “clip-like” perception and the absence of details of incidents in memory. (Incidentally, this is anomalousness. Human consciousness is constantly looking for familiar images and sometimes finds them. Recall at least the sphinx from Mars)
3. Since the previous two points are still not enough, the brain begins to extrapolate (predict) the possible development of events. This is especially noticeable in athletes, such as volleyball players, who hit the ball not because they had time, but because they guessed where it would be.
Extrapolation systems (including humans) have specific disadvantages. Such systems err from time to time. For a bipedal andoroid, this is not critical, you think, it will slightly falter, but for a robot controlling a nuclear reactor it is not excusable. What is permitted to Jupiter (man) is not allowed to a bull (automaton).
Artificial intelligence systems should become a fusion of a deterministic (based on the “just arrived” information) automaton with an extrapolation (able to “predict” based on past information). The predictive part should work in rapidly changing non-linear ways (how to score a goal), and the deterministic part should solve problems that do not require an immediate answer (what to wear for breakfast). The task of creating an extrapolation automaton does not seem trivial, and its solution is seen only through self-training. But nature did the same. Why are we worse. Look, the robomules are already jumping all over the stones, and, at the same time, they could also sing songs of their own composition. And you need to create such an automaton only once. For specific tasks, he will learn himself and much faster than a person.
Currently, the problem of artificial intelligence is usually reduced to modeling neural connections by program methods or to algorithmizing all possible situations and answers to them. These are attempts, at one of the levels, to copy the natural brain. In the first case, at the elemental (neurons and connections between them) level. In the second case, on the instinctive (stimulus-reaction; “if - then” construct). Both for the first and the second path, large computational powers are required, which, according to some, can only be given by quantum computers of the future.
It seems to me that the problem is not only in power. The “root of all evils” of this approach is that the developed artificial intelligence systems are “real-time” systems. That is, the stimulus - data processing - is the answer. It seems the animal’s brain works the same way, but it’s not quite so. The brain of a sufficiently developed animal (we will not take worms and lancelets) does not work in "real time", but in the "extrapolating" mode. As a reaction, a certain pre-calculated answer is issued, not an instant reaction.
This is due to both the relatively small processing power of the brain and the large amount of sensory information coming into it. For example, take the skin of a mammal. From each hair comes a nerve, that is, a source of information, and how many of those hairs (in humans from 200,000 to 1,000,000). And everyone feels not only the fact of touch, but also the power of this touch.
In addition, from the moment of exposure to the sensory organs to the response of the muscles, a lot of time passes (tenths of a second) during which the body decides to live (to catch prey) or die (to miss prey). Our brain (as well as the brain of other higher animals) is forced to live in the past.
To circumvent these problems, evolution has found the following solutions:
1. Signals from the senses are not immediately transmitted to consciousness. The brain is forced to filter them at the subconscious level, thus bringing to consciousness only what is “important”.
2. Pattern recognition is carried out at the "screen" level. That is, slightly similar - it means it. To further accelerate the response of the body, information is not seen in a detailed picture, but discrete, simplified images. Moreover, the degree of simplification increases in stressful situations. In humans, this is expressed in the “clip-like” perception and the absence of details of incidents in memory. (Incidentally, this is anomalousness. Human consciousness is constantly looking for familiar images and sometimes finds them. Recall at least the sphinx from Mars)
3. Since the previous two points are still not enough, the brain begins to extrapolate (predict) the possible development of events. This is especially noticeable in athletes, such as volleyball players, who hit the ball not because they had time, but because they guessed where it would be.
Extrapolation systems (including humans) have specific disadvantages. Such systems err from time to time. For a bipedal andoroid, this is not critical, you think, it will slightly falter, but for a robot controlling a nuclear reactor it is not excusable. What is permitted to Jupiter (man) is not allowed to a bull (automaton).
Artificial intelligence systems should become a fusion of a deterministic (based on the “just arrived” information) automaton with an extrapolation (able to “predict” based on past information). The predictive part should work in rapidly changing non-linear ways (how to score a goal), and the deterministic part should solve problems that do not require an immediate answer (what to wear for breakfast). The task of creating an extrapolation automaton does not seem trivial, and its solution is seen only through self-training. But nature did the same. Why are we worse. Look, the robomules are already jumping all over the stones, and, at the same time, they could also sing songs of their own composition. And you need to create such an automaton only once. For specific tasks, he will learn himself and much faster than a person.