How intelligence works (a single algorithm for discernment and generalization)

Do you know exactly how you know something? Nobody knows!

I want to reveal to you several secrets of natural intelligence, and at the same time tell you how to design artificial intelligence.

A small disclaimer. The article will describe very ambitious ideas. Most of the ideas presented can be expanded into independent cycles of articles. Therefore, ideas are presented here only for the initial acquaintance. I have no illusions that there will be many who grab the idea of ​​rally. Therefore, questions are welcome, I will try to clarify. And yes, I know that this is all very similar to a million other ideas, algorithms, etc. The only difference is that this design of ideas claims to be a completely complete simulation of the work of natural intelligence in all aspects that you can or cannot imagine . No black holes, unreasonable problems or unknown technical solutions.
')
Yes, it's all about the most complete and powerful artificial intelligence. I anticipate the greed of some researchers and the neglect of some collectors of ideas. But nevertheless, now, after the warning, we will start. There will be a lot of information and its density is very high, so - hold on to something stronger. It may be necessary to re-read dozens of times and ask thousands of questions. I am ready to answer them, since it is time to bring the single studies to a more practical level, requiring the involvement of several hundred specialists.

For a start, let's see how the mind works when it finds out. It would seem that we are so quickly guided in the setting, just for a few moments, to find out what objects we see, how they are located, what their behavior is. Because of this illusion, it seems that the brain works very, very quickly.

Why is this an illusion? I'll try to explain.

Imagine that you unexpectedly fell into a very unfamiliar situation. There, where they have never been. For example, in the jungle, or vice versa, in the desert. Or maybe it will be just a thick ragged fog? Or you see someone else's starry sky.

What's happening? What distinguishes the mind? He has nothing to rely on. It is necessary to classify the decor slowly and painfully. Seek out reference signs clusters in order to navigate relative to them.

Do you know what the newborn sees? Chaotic movement of colored spots without content and meaning. But something he hears that allows him to start navigating. He hears the familiar voice of the mother. He was already accustomed to touch her body, warm, fragrant, giving delicious milk.

The process of thinking is laid starting with the formation of the sense organs.

When the mind finds itself in an unfamiliar situation, it does not understand it. He has nothing to rely on. There is nothing that he can predict.
And he begins a consistent and deep classification of the situation.

In order to somehow move from images to algorithms and mathematics, we need only four terms and the relations between them.

The first term is a sign. A sign is something that can be perceived. For example, black. Or salty taste. Or round shape.
The second sign is the presentation. The presentation is based on the recurrence of symptoms. If two signs are perceived together, a representation arises that summarizes these signs.
The third term is a subview. A subrepresentation is a relation of a feature to a presentation. For all views, its attributes are subviews.
The fourth term is a super sign. The super sign is the relation of the view to the super sign. For all the signs of the presentation in which they are included - these are super signs.

All together, this is combined into a proactive universal classifier.

What does it mean - proactive? This means that he classifies all the time. This is extremely important. All behavior of the classifier is defined only by where it wants to look, what to check.

What does universal mean? This means that he classifies everything he can reach.

What does the classifier mean? This means that he builds a connected network of ideas about situations, about trends in their changes, about the most stable ideas of the situation, thereby forming the most stable idea about the external world.

And how does all this work?

The classifier algorithm is not complicated, but not quite easy to understand. First I will describe the requirements for the algorithm, and then I will give a couple of illustrative examples.

1. The organs of perception are represented by specific drivers for each sphere of perception.
1a No matter how complex the sphere of perception is, it unfolds into a one-dimensional set, in which the attribute corresponds to the address (index) and the value of the attribute on this index
1b. The classifier, referring to the driver at a specific address, receives the measured value as a result. This can be a character in a sequential text, the color of a region in space, the height and amplitude of a sound, temperature, etc.
1c. It is important that the driver uses a relative addressing model, from index to index, without in any way preserving the time sequence. For example, in the line “MOTHER SOAP FRAME” the index denotes the offset. If the current index indicates the first letter of the second word, then the offset +2 indicates the letter "L". If you again set the offset to +2, we get a space.

2. The classifier saves the transitions, counting the probability of confirming and disproving the values ​​during the transitions from the key feature by the index offset.
2a If the current value of the feature is already in the network, then the classifier chooses the most likely representation for this feature and predicts the most likely transition to the next feature, after which it calls the driver with an offset.
2b. The entire network of representations is constantly adjusted and sorted according to the most probable ideas about current features.
2c. Representations of the second order connect among themselves representations of the first order. Thus, in order to verify the representation of the second order, it is necessary to verify the representations of the first order contained in it. It turns out a recursive algorithm.

3. The instantaneous state of a classifier is represented by a hierarchy of ordered lists of representations of different orders.
3a The algorithm always tries to increase the order of presentation in the first place. The highest order of representation reflects the highest level of understanding. So, for example, for the lid and legs, the next level of presentation is the table, and for the table, chair and sink, the next level is the kitchen.
3b. The classifier contains simple numeric numbers as representations. This is the internal language of the classifier. The internal language in a subset corresponds to the language used by people for communication (expression of representations and their recognition). The algorithm for comparing representations expressed and perceived is exactly the same, but hack is also possible. You can choose a specific order of representations and increase the priority of those that make sense to a person. Then the classifier will quickly explore precisely those ideas about the world that will help it to more quickly and better navigate the world perceived by people and communicate with us in the same language.

4. The classifier does not work with infinite memory and in infinite time. Therefore, he ignores quite a few signs and ideas as unpromising for a stable classification. Later we will look at how multidimensional spheres of perception, including movement and measure of movement - time, are packed into a network of representations.
4a. Previously, a million transitions between the subrepresentations of the previous order are assigned to each order of representations, while for the formation of the next order only twenty thousand of the most probable are taken into account. It is important here that both the most probable confirmations and the most probable refutations are useful, therefore we consider them separately.
4b. The number of orders of representations is large enough, and almost equal to the number of representations of the first order. In order to reduce the number of orders (after all, for each order we allocate about a million transitions between subviews) from one million to the same 20 thousand, the idea of ​​the most upper context is used. Submissions confirming the presentation of the highest order after ranking by probabilities are thinned out, retaining only the shortest chains to the primary signs.

5. The direction of classification is always determined from the presentation of the highest order. The context contains the highest order representations ordered by probability, constantly updating this context, checking for the occurrence and termination of signs, most quickly confirming or refuting this context.
5a. The content of this context is the most complete understanding of the situation studied and trends.
5 B. The smaller in the context of understanding the representations, the higher the assessment of clarity and accuracy of understanding
5c. It is within the context of the context that the issues of communication, the distinction between subjects and objects, their groups and interactions, goals and desires, relationships and emotions are determined. It is possible to consider this question for a long time, for a start there should be a good idea about the main cycle of the classifier and the main slice of the state of the classification network.

Why all these details?

Of course, there is a simpler idea of ​​the main idea of ​​the classifier’s work. This idea is: “what does it belong to”? This can be difficult to understand. When we talk about a cognitive act, we usually mean the question "what is it?"
Practically, this question has two aspects: “what does it own?” And “what does it belong to?”. These two questions are invariant, it is enough just to expand the direction of the search. These are parallel processes of perception / behavior. Perception moves in the direction of searching for the owner (the most stable support for further classification), and behavior moves in the direction of searching for the owner.

What next?

For those who are interested, we can join forces in the study of the model and the development of useful and relevant applications that do not require significant resources, but work with the most important in the AI ​​industry - with understanding.
Few places you will find something specific on the subject of understanding. What is understanding? How to define it? How to model it?

Therefore, I propose a very deeply developed model, covering both metaphysical and philosophical issues, and pragmatic technical issues of product implementation and implementation.

Welcome to the new world - the world of real artificial intelligence and artificial personalities.

Your move, habrazhiteli.