Shortcut to Artificial Intelligence?
Now we have many algorithms that perform certain (separate) cognitive functions. Some beat us to games, others drive cars, still others... I don’t tell you. We have created computer vision programs that distinguish road signs better than ourselves. Programs that draw and write music. Algorithms make medical diagnoses. Algorithms can put us in a belt in the recognition of cats, but... specifically this one, which for cats, is nothing but the recognition of cats. And we want a program that will solve any problems! We need a “strong” or “universal AI”, but without our own consciousness, so that we can’t refuse to solve the task, right? Where do we get it?
To understand how intelligence works, we turn to the only example we have. To the human brain, in which, as we believe, the intellect "lives". Someone will object - many living things have brains! Let's start with the worms? It is possible with worms too, but we need an algorithm that solves not human problems, but our human tasks, right?
Our brain. Imagine it. Two kilos (to the maximum) of a supple pinkish-gray substance. One hundred billion (we also take to the maximum) neurons, each of which is ready to grow up to ten thousand dynamic connections - synapses that can either appear or disappear. Plus, there are several types of signals between them, and even a glance threw a surprise - it also conducts something, helps and helps. (For reference: neuroglia or just glia is the aggregate of auxiliary cells of the nervous tissue. It makes up about 40% of the central nervous system. The number of glial cells is on average 10-50 times more than neurons). Dendrites recently surprised - it turns out that they perform much more functions than previously thought (1). The brain is a very complicated thing. If you do not believe it, ask Konstantin Anokhin. He will confirm.
Man does everything with the help of the brain. Actually, we - this is it. Hence, it’s completely unsurprising that a person has the idea that “brain=intelligence” and even more unsurprising is the idea to copy the device of the brain and - voila! - get the search. But the brain is not intelligence. The brain is the carrier. "Iron". And Intellect is an algorithm, "software." Attempts to repeat software through copying iron is a disastrous idea. This is the cult of cargo (2). Do you know what the Cargo Cult is?
The natives of the islands of Melanesia (seeing during WWII how planes bring weapons, food, medicine and much more) made straw copies of planes and a dispatcher’s booth, but they didn’t help themselves in getting the goods, because they had no idea that hiding behind the appearance of airplanes. So we, having analyzed the calculator to the cogs, will not find a single digit inside. And, moreover, no hint of any operations with numbers.
A couple of years ago, Andrei Konstantinov in one of the issues of the magazine Schrödinger’s Cat (No. 1-2 for 2017), in his column “Where the Robot has a Soul”, wrote: “Since Leibniz, we have not found anything in the brain, except for "parts pushing one another." Of course not found! And we will not find. Using computer hardware, we are trying to restore the program, but this is impossible. As a supporting argument, I will quote a long quote (3):
"... neuroscientists, armed with methods commonly used to study living neurostructures, tried to use these methods to understand how a simple microprocessor system functions. The “brain” was MOS 6502 - one of the most popular microprocessors of all time: an 8-bit chip used in many early personal computers and game consoles, including Apple, Commodore, Atari. Naturally, we know everything about this chip - after all, it was created by man! But the researchers pretended not to know anything - and tried to understand his work by studying the same methods that study the living brain.
The lid was chemically removed, under an optical microscope, the circuit was studied up to an individual transistor, a digital model was created (here I simplify a little, but the essence is correct), and the model is so accurate that it was possible to run old games on it (Space Invaders, Donkey Kong, Pitfall). And then the chip (more precisely, its model) was subjected to thousands of measurements at the same time: during the execution of games, the voltages on each wiring were measured and the state of each transistor was determined. This generated a data stream of one and a half gigabytes per second - which has already been analyzed. Charts of bursts from individual transistors were built, rhythms were detected, elements of the circuit were found, the disconnection of which made it inoperative, there were mutual dependencies of the elements and blocks, etc.
How complex was this system compared to living ones? The 6502 processor, of course, doesn’t even have a mouse near the brain. But it comes closer in complexity to the Caenorhabditis elegans worm - the biologists’s crowhorn: this worm has been studied along and across and attempts are already being made to simulate it completely in digital form (...) Thus, the task of analyzing a system on a 6502 chip is not an over-simplification. And the results are eligible to be extrapolated to in vivo systems.
That's just the researchers... were defeated! No, some results, of course, were obtained. By analyzing the chip, we managed to isolate the functional blocks, sketch out a diagram of their probable relationships, and get some interesting hints about how the microprocessor as a whole probably works. However, understanding in the sense in which neurobiology requires it (in this case: to be able to correct any breakdown) was not achieved. "
At some point, researchers appeared who began to say roughly the same thing - that you need to study algorithms, that you need to understand what function the intellect performs. For example, Demis Hassabis (DeepMind), preparing for a speech at the Singularity Summit in San Francisco (2010), said the following: “Unlike other speeches at the AGI Summit, my report will be different because I'm interested the systemic level of neuroscience - the algorithms of the brain - and not the details of how they are realized by the brain tissue in the form of spikes of neurons and synapses or specific neurochemistry, etc. I am interested in what algorithms the brain uses to solve problems, and which we need to find in order to get to AGI. "
However, after 10 (!!!!!) years, everything continues as before: scientists examine the brain and try to figure out how the process of interest occurs from the external manifestations of physiological activity and its internal structure. How many tasks - so many processes. People are all different. Everyone’s brains are a bit different. Of course, there is some kind of averaged picture, however... Imagine that at any arbitrary moment in time the brain solves a lot, including “subconscious” tasks, monitors and controls the internal state of the body, perceives and interprets the signals of the external environment (and we don’t talking about numerous feedback loops). Can we confidently identify, reliably identify and clearly separate these “activities” from one another? Is this possible in principle? Honestly, I doubt it. Not to mention the reproducibility of these processes on non-biological media...
Let's look at the situation differently. What is a “task” in general? This is a difficult situation that one faces and which a person is trying to resolve. As American mathematicians Herbert Simon and Allen Newell showed in the middle of the last century, any problem in its general form can be described as a transition from the “System with a problem” state to the “System without a problem” state. They developed a computer program, calling it “General Problem Solver” (Universal Problem Solver), but they did not advance further than solving specific problems, so the universality of their algorithm remained in question. But the "System with Problem" formula is > “System without a problem” turned out to be absolutely true!
The transformation of the System is the process of its transfer from the initial state “with a problem” to the desired state “without a problem” (4). In the process of conversion, (i.e.solutions to the problem) the problematic system becomes problem-free (well, or less problematic), improves, gets rid of its shortcomings and “survives”, that is, continues to be used. Oh wait, what did we just say? Getting rid of the flaws? Survival? Hmm... Something familiar. Somewhere we are... Ah, yes. Evolution! The fewer flaws - the greater the chance of survival!
Let's check ourselves, remember and repeat the main postulate: in wildlife, organisms that have a greater number of useful properties have a better chance of survival (well, conditionally - the horns are branched, the tail is more magnificent). If the body has paler feathers and a nastier voice (harmful properties), then, most likely, his life will be short-lived and pass by himself. As a result, the pressure of selection leads to the fact that organisms get rid of deficiencies and become more and more viable. If you do not believe it, ask Sir Charles Darwin. He will confirm.
So, we accept as a fact that
a) the function of intelligence is the solution of problems (any) and
b) the solution to the problem is the improvement of the System (any), during which it gets rid of the shortcomings, becomes more viable. In other words, it is evolving.
Hear the crack These are our ideas about the complexity of the intellect that are beginning to burst at the seams. It turns out that the previously existing concepts of “brain complexity” and “intelligence complexity” cease to be identical. What if for “obtaining Intelligence” it is not necessary to carry out “reverse engineering” of the neurophysiological process of solving the problem, catching the ghostly shadows of thinking in a connection (especially since each person is unique) or engage in deep learning networks? What if... we need to algorithmize the process of evolution of the system, that is, the way of its transformation from a less perfect state to a more perfect one using the laws of evolution known to us? What if until today we really decided the wrong task?
However, I do not want to say that training networks is not necessary. This and other areas have great prospects. And even more so, I do not want to say that deep research into the physiology of the brain is a waste of time. Studying the brain is an important and necessary task: we better understand how the brain works, learn how to treat it, recover from injuries and do other amazing things, but we won’t come to intelligence.
Someone will object to me now: the tasks that a person solves are connected with millions of different systems - natural, social, industrial, technical... Material and abstract, located at different levels of the hierarchy. And they each de develop in their own way, and Darwinian evolution is about wildlife. Bunnies, flowers, fish, birds... But studies show that the laws of evolution are universal.
Evidence is not long to look for - they are all before our eyes. Let those who have them see. Whatever you take - from a match to a Boeing, from a tank to... double bass - everywhere (5) we see heredity, variability and selection! And the whole variety of evolutionary changes (the apparent complexity of which is due to the fact that all systems are very different in nature and are at different levels of the hierarchy) can be expressed in a single cycle. Do you remember, huh? "System with a Problem" - > "System without a problem."
What is a system with a problem? This is a System (material and abstract, social, industrial and technical, scientific and... any - an object, idea, hypothesis - anything), in which some flaws are found that affect (attention!) Our desire and the possibility of its use. The system is not good enough. The system is not efficient enough. She has a low benefit/cost ratio. We want, we can and are ready to refuse it, and we often refuse. But we need another (performing the useful function we need), but already “no problem” - more efficient, without flaws (or with fewer of them). Well, you saw this picture above... Of course, one “arrow” between the two extreme states (initial and desired) is not enough for us.We need the same “operator", "converter", right? Let's try to find him? You will agree that in case of success we will receive a description (at least for a start and simplified) of the universal algorithm we need so much?
Starting point - “System with a problem”. We are starting to think about how to abandon its use. The moment we call (or feel) “We have to do something!”
The reason that threatens the survival of the system is low ideality, which is expressed in a reduced value of the ratio of useful functions of the system to costly (harmful) functions.
What do we do next? We either a) create a new system (if the system with the necessary functions either does not exist, or the existing system does not have resources for improvement) or b) we improve, modify, the existing one (if there are still resources). We study the internal structure and deal with the external environment - we identify the external and internal shortcomings of the System and after their elimination we get an improved System. A system with enhanced ideality and increased vitality!
Due to the fact that the above Scheme describes the process of development, improvement, or, if you want, evolution of any systems (which is easy to verify by substituting any other optionally from “Lampshade” to “Anchor” instead of the word “System”), I I think it can be safely... and even - it is necessary! call the "Universal Evolution Scheme." And pay attention - it is absolutely algorithmic, that is, it fully falls within the definition of an algorithm: Algorithm is an exact order to execute in a certain order a certain system of operations leading to the solution of all problems of this type. then it can be implemented as a computer program).
In the presented form, the Universal Evolution Scheme:
- natural - the laws of evolution are revealed in systems of various types, and their action is tested in technology, production, society, nature and thinking;
- objective - the laws of evolution do not depend on the opinion of the researcher and/or user;
- logical and consistent - the laws of evolution follow one from another;
- complete - a set of laws of evolution is sufficient to describe any system;
- rigid - the laws of evolution cannot be rearranged and
- closed - the laws of evolution form a cycle: the system, having passed one cycle of changes, immediately begins a new one.
What we get as a result: The evolution of the system (presented in the form of a Universal Scheme) is a way to improve it, to get rid of shortcomings. In other words, this is an algorithm for solving the problem. And solving a problem is exactly what the intellect does. Simplify and get: Universal Scheme=description of the function of intelligence.
Constructive criticism is welcome.
1. “Dendrites are more important for the brain than previously thought” chrdk.ru/news/dendrity-vazhnee-chem -schitalos
3. E. Zolotov. "Understand me! How the inanimate helps to understand the living » www.computerra.ru/161756/6502
4. Chapter 6. Problem Solving. Artificial Intelligence. A Knowledge-Based Approach by Morris W. Firebaugh University of Wisconsin - Parkside PWS-Kent Publishing Company Boston 1988, p. 172.
5. Darwin evolution in the world of the technosphere. The world of things created by man develops according to the same laws as wildlife. www.ng.ru/science/2017-01-11/14_6899_evolution.html .