What could be the AI program
Below, I wrote a short discussion about what an AI computer program could be. First, in my opinion, it is necessary to separate the mind itself, which answers the question “HOW to reach the goal?”, From the desires that define the goal, in other words, the mind solves the tasks set by the emotions. Further, the mind means a program or algorithm that can find answers to questions when the answers are not known in advance. But how could this program work, how does it figure out the sequence of actions necessary to achieve the goal?
Here it should be immediately noted that in order to work out the simplest algorithm for achieving the goal, the mind must be able to identify the connections between events, and automatically. Events are understood as some abrupt changes in the surrounding world, for example, the circle broke or the sound sounded.
In nature, not the smartest animals have this ability: for example, a cat establishes a connection between event 1 (nickname) and event 2 (feeding). But how does this happen? And what events to record in memory and what not? After all, a lot of events, and they all obviously do not fit in the memory. For living beings, connections are apparently established in two cases: either with repeated repetition of events, or if an event causes a strong emotion. For AI, in principle, you can set a variety of criteria.
How could the formation of connections between events, objects, in a computer program be realized? If, for example, a program receives a video stream and an audio stream, then these streams should be represented as a series of events. For example, a video camera sees how the green lamp lights up first, then red, then green again, etc. This should be displayed in the sequence of events ---- ... etc. In this situation the memory should be fixed that event K is followed by event W, and after event W is followed by K. Or an example: the video camera sees how the lamp lights up when the switch is turned up, and when it moves down it goes out. The memory should be fixed, that the movement of the knife switch up turns on, and down turns off.
')
But what does this mean for a computer program? And the fact that in the section of memory in which the event of “raising the switch” is stored, there is a pointer to another event, namely, to turn on the lamp. Etc. Here the question arises: how to separate events from each other in a continuous flow of information? It can be assumed that simply subtracting the previous picture from the current one, for example, pixel-by-pixel, and if there is a difference, say, 10%, then take it as an event. Although the criteria may be different.
Now about the connection between objects: when a person hears something or sees something, then he has various associations. That is, other similar images, as well as sounds, objects that are close in context are associated with the image of the object that the person sees. For example, the word tea is associated with sugar, a mug, a spoon, etc. And here is the question: how could this be done in a computer program? Apparently, the very image of the object or its sound should be like its name in memory. That is, when the mind sees an object, then its image is the name of the “folder” with information about it. Moreover, the folder may contain references to other images, events and information associated with the current object.
It should also be possible to store in memory not only specific images, sounds, events, but also some generalized formulas, such as a triangle, motion, and so on. For example: there can be infinitely many different triangles, but all of them have a definite formula that describes what a triangle is. How to implement it?
Everything is simple here: on one side, images of different triangles are fed to one sensor, and the same designation is on the other. That is, the same signs of different triangles should be preserved in the memory cell with the name "triangle".
Of course, the mind operates with information about the world around it stored in memory, but it is very compressed and most likely reworked in a certain way. That is, if pictures are saved in memory, then this is definitely not a BMP format, but something closer to SVG. And if the sound is not WAVE. And if these are processes, events, then this is not a video, but rather a short sequence of SVG pictures in a sense, like gifka, although the trajectory of changes is also possible.
Now the question arises: how should all this work? Reason must, after all, UNDERSTAND the question that is asked of it. But what does it mean to understand the question? Suppose a question is written in human language. Then the AI must understand what the words and phrases of which the question consists. But how is it: to understand words? How can a computer program understand?
In case the word means an object, then this word for AI should indicate information about this object (image, some sounds, etc.). For example, the word guitar should indicate the shape of the guitar, and the sounds that it makes.
However, many words, such as “movement”, “add”, etc., denote actions, processes. What information should they point to? Here we need the ability to generalize. Then the movement will be associated with the word: changing the coordinates in the space of something. Etc. Thus, the ability to establish relationships between events, objects will cause the AI to understand words.
Suppose an example: ask the AI the question: “how to make the liquid sweeter?” It is obvious to us: Make it sweeter = increase the parameter “sweetness”. In order for an AI to give an answer, the following interrelations should be in the AI memory: the sweetness of the liquid is directly related to the amount of the sweet substance in it. And the phrase “make sweeter” meant: increase the sweetness setting.
Or another question: how to make the sound of the receiver louder? For us, the answer is obvious: scroll the receiver knob, for example, to the right. In order for the AI to give an answer, it is necessary to have interrelations in its memory: the volume depends on the position of the knob, rotation to the right = increase, make louder = increase the volume.
So, the AI program should be able to do something like this:
- Convert information from sensors into a simplified form in which it is stored in memory;
- The program should establish the simplest causal relationships between events, the cause and effect are determined by their sequence;
- The program should be able to make generalizations, that is, to identify the common features of various objects, phenomena, events;
The memory should be organized so that the image, sound, “opens a folder” with information about this object and all other objects like it and associated with it.
Here it should be immediately noted that in order to work out the simplest algorithm for achieving the goal, the mind must be able to identify the connections between events, and automatically. Events are understood as some abrupt changes in the surrounding world, for example, the circle broke or the sound sounded.
In nature, not the smartest animals have this ability: for example, a cat establishes a connection between event 1 (nickname) and event 2 (feeding). But how does this happen? And what events to record in memory and what not? After all, a lot of events, and they all obviously do not fit in the memory. For living beings, connections are apparently established in two cases: either with repeated repetition of events, or if an event causes a strong emotion. For AI, in principle, you can set a variety of criteria.
How could the formation of connections between events, objects, in a computer program be realized? If, for example, a program receives a video stream and an audio stream, then these streams should be represented as a series of events. For example, a video camera sees how the green lamp lights up first, then red, then green again, etc. This should be displayed in the sequence of events ---- ... etc. In this situation the memory should be fixed that event K is followed by event W, and after event W is followed by K. Or an example: the video camera sees how the lamp lights up when the switch is turned up, and when it moves down it goes out. The memory should be fixed, that the movement of the knife switch up turns on, and down turns off.
')
But what does this mean for a computer program? And the fact that in the section of memory in which the event of “raising the switch” is stored, there is a pointer to another event, namely, to turn on the lamp. Etc. Here the question arises: how to separate events from each other in a continuous flow of information? It can be assumed that simply subtracting the previous picture from the current one, for example, pixel-by-pixel, and if there is a difference, say, 10%, then take it as an event. Although the criteria may be different.
Now about the connection between objects: when a person hears something or sees something, then he has various associations. That is, other similar images, as well as sounds, objects that are close in context are associated with the image of the object that the person sees. For example, the word tea is associated with sugar, a mug, a spoon, etc. And here is the question: how could this be done in a computer program? Apparently, the very image of the object or its sound should be like its name in memory. That is, when the mind sees an object, then its image is the name of the “folder” with information about it. Moreover, the folder may contain references to other images, events and information associated with the current object.
It should also be possible to store in memory not only specific images, sounds, events, but also some generalized formulas, such as a triangle, motion, and so on. For example: there can be infinitely many different triangles, but all of them have a definite formula that describes what a triangle is. How to implement it?
Everything is simple here: on one side, images of different triangles are fed to one sensor, and the same designation is on the other. That is, the same signs of different triangles should be preserved in the memory cell with the name "triangle".
Of course, the mind operates with information about the world around it stored in memory, but it is very compressed and most likely reworked in a certain way. That is, if pictures are saved in memory, then this is definitely not a BMP format, but something closer to SVG. And if the sound is not WAVE. And if these are processes, events, then this is not a video, but rather a short sequence of SVG pictures in a sense, like gifka, although the trajectory of changes is also possible.
Now the question arises: how should all this work? Reason must, after all, UNDERSTAND the question that is asked of it. But what does it mean to understand the question? Suppose a question is written in human language. Then the AI must understand what the words and phrases of which the question consists. But how is it: to understand words? How can a computer program understand?
In case the word means an object, then this word for AI should indicate information about this object (image, some sounds, etc.). For example, the word guitar should indicate the shape of the guitar, and the sounds that it makes.
However, many words, such as “movement”, “add”, etc., denote actions, processes. What information should they point to? Here we need the ability to generalize. Then the movement will be associated with the word: changing the coordinates in the space of something. Etc. Thus, the ability to establish relationships between events, objects will cause the AI to understand words.
Suppose an example: ask the AI the question: “how to make the liquid sweeter?” It is obvious to us: Make it sweeter = increase the parameter “sweetness”. In order for an AI to give an answer, the following interrelations should be in the AI memory: the sweetness of the liquid is directly related to the amount of the sweet substance in it. And the phrase “make sweeter” meant: increase the sweetness setting.
Or another question: how to make the sound of the receiver louder? For us, the answer is obvious: scroll the receiver knob, for example, to the right. In order for the AI to give an answer, it is necessary to have interrelations in its memory: the volume depends on the position of the knob, rotation to the right = increase, make louder = increase the volume.
So, the AI program should be able to do something like this:
- Convert information from sensors into a simplified form in which it is stored in memory;
- The program should establish the simplest causal relationships between events, the cause and effect are determined by their sequence;
- The program should be able to make generalizations, that is, to identify the common features of various objects, phenomena, events;
The memory should be organized so that the image, sound, “opens a folder” with information about this object and all other objects like it and associated with it.
Source: https://habr.com/ru/post/237631/
All Articles