Summary of the problem of “two or more teachers” and a subjective opinion about the AI community
While I was here I was thinking the idea was spreading through the articles
1. Model of functional separation of consciousness and unconscious. Introduction
2. Model of the manifestation of consciousness or ANN without the effect of forgetting
3. The problem of "two or more teachers." First touches
4. Training with reinforcements on neural networks. Theory
The audience seemed to have the necessary knowledge, all the same, we have the color of society - programmers :) But alas ... the last survey showed that not all programmers are aware of AI issues. And boorish students who have run into these articles in the comments have not yet completed their studies.
')
Let's try to summarize and consider this extended article promised in the survey. And at the same time I was told that there are serious problems in the AI community. After a number of comments - yes, apparently there really is. Let's try to look at the trend.
Let's start with this. Many people will not like it, but try to soberly evaluate it yourself.
As in any science, AI has already acquired a full-fledged infrastructure :). And this is what it looks like:
1. At the foot of the AI pyramid, oracles sit - these are the people who loudly declare that we are engaged in "strong AI", "weak AI" - we are not interested. These freaks - at least the ones I met were all freaks. These are creationists of the 21st century, they adopt logic and try to build fictitious theories. No more no less. If you are a young man who is still not tied up - get out of this swamp, it sucks good smart heads. At first, do not even talk to such people - they feed on your energy, like religious sects.
2. Engineers are located on the steps to the AI pyramid. They do not stop to engage in "weak AI", often it is their bread. Even at the institute, they choose a very narrow direction - recognition of machine numbers, image processing, stock prediction, robot control, etc. etc. Often they have the title of doctor of science, and a little less often have the title of professors. If you are a young man who got to work with such engineers, or listen to their lecture - soak up the basics, otherwise you will certainly then slide to the foot. But having absorbed the basics - do not get carried away. Do not become a student who knows only the terminology - express your thoughts naturally. Try to do something yourself, never judge an algorithm, an approach only on the basis of what such an engineer said to you. It often turns out that he just read a lot of books and judges them (yes, I knew good AI professors who admitted this, which only did them more honor). If you believe him in what he does not directly deal with, then without checking the originals, in a couple of years you will transfer this knowledge from hearsay. As a result, you'll make a little more than a mistake. Of course, to check the originals and try to reproduce the experiments of the authors of the methods is time consuming. But otherwise, over time, you will have to believe a false impression that you understand something in the neighboring areas. And you will no longer become just a cheeky student who knows only terminology, but a person with authority, because you were engaged in a private (you will proudly call her a specific and concrete result) task, but at the same time you will distribute obviously inaccurate impressions that did not try experimentally.
3. On the steps in the shade, a little away from the heat, the neurophysiology engineers crouched, they got so tired of doing their work with patients, that to think about the other side, the AI engineers shout - no longer possible. They quietly take and do in their experiments, something that can be disassembled in the distant hum of the hassle of AI engineers. Something that they have, that there is no ... but as a rule they do not keep up with the thought of AI-engineers. But due to the fact that they really embody it in a human-machine interface, it always looks impressive.
4. But the question is - who is located in the gazebo on the top of the pyramid? It is not strange - it is empty, no input is free, just no one goes there. Sometimes only fleeting engineers in the heat of a fight fly in there, and also quickly leave. This is a place for AI theorists. Engineers often do not distinguish where the bottom, and where up - so they confuse theorist with the oracle. But this is how the brain of an AI engineer works; he, seeing theoretical generalizations, doesn’t see a specific task and doesn’t understand if it’s a bug. He does not see the forest for the trees, and seeing the sun, he takes it for the revelation of the oracle - and he thinks, ugh, why should I go up to the sun to do something not concrete. So he does not rise above the trees, fearing to be burned by the sun. And those who burned - he sees - groveling under his feet, on the outskirts of the pyramid. But of course, if he studied from the words of his contemporaries - he doesn’t know that those who created the concept of neural networks, genetic algorithms, fuzzy logic and much more ... still visited this gazebo at the top of the pyramid, just not very accepted speak.
When at last it happens to push through a crowd of AI engineers and an unenviable fate awaits an AI theoretician in the arbor. From above it becomes clearly visible that he does not have the right for simple solutions to simple problems. He can no longer afford to take a private, but previously favorite, specific task. He sees that in order not to be pushed out of the gazebo, he needs to cool down a little the heat of the crowd on the steps of the pyramid. And you need to splash water on everyone and not on any separate group.
But how can you convince that the splashed water is suitable for drinking by AI-engineers? After all, they only need gadgets to solve specific problems. Here the most difficult. It is necessary to degenerate the tasks of AI-engineers, only then they will be something common. And find a solution to this degenerate problem. Further, in fact, it is not lordly to adapt degenerate tasks to the real world from which AI engineers derive their tasks. They need to be given at least a tool, and already their problem is to use a hammer or an ax.
But it is becoming more and more difficult for the modern theorist, unpretentious degenerations have already been built - there are already ideas about classification problems, clustering problems, MDP, PoMDP, etc. We have to degenerate tasks so that they would increasingly remain the problems of the real world.
So we consider one such degenerate problem.
At first I’ll have to take off my rose-colored glasses from those who wear them. Let's talk about the ability of the INS to predict and the speed of their learning. Everything is in accordance with Minsky - they forecast badly and slowly. And although I will be showing on the Rosenblatt perceptron, there is no fundamental difference for other networks.
This is for you AI engineers select a beautiful sample of tasks with good results on specific tasks. I'll show you the situation in the forehead. Who does not agree - we take the task described below and test it ourselves and provide specific results, with the code laid out - so that you can reproduce the results (I am tired of boltology - have entered).
Task.
We take a simple function c = a + b, where a and b are integers from 1 to 64, and c is an integer from 1 to 256 (we will need a slightly larger dimension later).
Then we will have 16 inputs and 256 outputs in the perceptron. Those. At the input we will submit 2 bytes representing in binary form the numbers a and b. And in order for the perceptron to work a little better, the output will be interpreted as what the output number is and the number. To get rid of the ambiguity of the way out - we work on the principle of "the winner takes all", i.e. We consider only the output that is most active.
The idea of the accuracy of the prediction of the neural network
If we teach the perceptron throughout the specified space of the inputs of the outputs, then it will unmistakably learn how to fold. To display this graphically, we draw a “red square” :) where, starting from the upper left corner along the X axis, put the number a, along the Y axis b, and the sum c at the intersection is displayed with a dot, and in the gradation of red. We get the following:
But when everything is known there is no place for prediction. So let's remove every second point:
Now take any graphics editor and complete the missing points. Can you restore the "red square". I think that only one person can do it - yes, yes to Malevich :)
But the perceptron will make a mistake just like any other person at almost every point, but will retain only the general tendency. Here is what he will do:
Thus, we see when mathematical accuracy is needed - predictions are fortune telling on the coffee grounds. But as an indication of some tendency is quite suitable. Therefore, the perceptrons work with image processing, since Our eye can be deceived, and detailed accuracy is not needed there. But in machines where the arithmetic precision is needed, the perceptron will not work. The same property and our intuition is with you - it can be said the use of our graphic memory in analytics. That's the whole psychology.
Fitness function
It's actually easy. The above function c = a + b we use may well reflect suitability for some process. (In economics, they like to talk about the utility function, they have everything there that is not useful, they don’t have to spend money :), or they’re also talking about the evaluation function - these are also fans of prices :) Well, we don’t ask for natural processes - we judge the suitability of suitable conditions for our actions). Well, let us heat the water in the kettle with two sources, separately from which are the temperature sensors of each source. We need to determine when the water boils.
in fact, we solved this problem above. Above, we trained the perceptron to perform the fitness function c = a + b. It remains only to beautifully display the temperature above 100 degrees, let's do it in the gradation of green that's what we get.
And we see that in general, where the perceptron has not abandoned the forecast (black dots), it is not wrong in principle in the trend. AI engineers can play around with parameters, types of neural networks, etc. and get almost no mistaken picture. In general, there will be many mistakes - but the border will be found. And thus completely fuzzy device will be able to give guidance to the action quite clearly. This is used by those who teach agents to run on the basis of a particular fitness function.
Now you just could not understand how to set the fitness function on an artificial neural network;)
Why is it convenient to set the fitness function using an artificial neural network?
This question is somewhat more complicated. It would seem, why use such a network that is inaccurate and slow in learning , when you can easily set the fitness function with a mathematical formula? All the same, that not to use intuition, and to use the bases of analytics received at school.
The problem with the mathematical formula is only one - it is tough and uncompromising.
But in practice, we began to measure the temperature and in previously unknown points - we began to receive some kind of anomaly. It is not known why, but they started. In the region of still quite small temperature sources - the temperature became sharply above 100 degrees. Here’s how it might look:
And now let's imagine each dimension is given to us very dearly, it takes several days and $ 10,000 - and understand what the anomaly is, and at least in what area we really want. Our formula is irrevocably obsolete and not suitable.
And here nothing is better and did not figure out how to use the neural network, although it can be bad, it can give a forecast - at least like this:
It will be a little more difficult. Get ready to think :), I guess that the above described became clear even to the watchman in the market.
Suppose now we have two professors, one of them measured the heating of our water from two sources and received a layout (on the right), and the second layout (on the left):
Then one analyzed and built the theory that the temperature of the sources must be added to get the final temperature. And the second built another theory - it is enough to take only one of the temperatures of the source that is hotter.
And so they met, of course they quarreled, calling each other ignorant, deceivers, etc. and did not agree on anything, each went to his laboratory. Ordinary story is it? Then each of them created his own “School”. He began to teach students to his theories, and each boorish student of one school argued with the same boorish student of another school - after all, they had to shout over each other - so students win a place in the sun on the steps of our pyramid.
But it was clear to one dropout for a long time watching the struggle of these schools. But there is probably some factor that the professor does not take into account both school A and school professor B. He decides to tell them, but they closed in the study, and they are not allowed people on the street, afraid to let in some oracle. Therefore, our dropout has to wade through the students of professors, they of course each chuckled in his own way, twisted his finger at his temple, and one even strove to drop by the nose. Realizing that things will not be, having matured and pootbivy side of our hero retired to the gazebo at the top of the pyramid. There he realized that until he showed what the factor of the divergence between the two professors would be, he would not be listened to. And then he looked at the people who were seated and crowded on the steps. And I understood - not, since I had already climbed so high - why should I try on two professors, how many of them argue with each other and each couple has a particular problem - the uncertainty of the general factor of divergence of theories.
And he began to think about the problem in general categories, then to sprinkle all those sitting on the steps of the pyramid. Yes - he became infected with the spirit of the arbor, and I understood why many people do not want to go here. It really breaks a person.
And that's what he understood for now. Using the neural network as described above to identify where the anomaly is fails. Anomalies are everywhere throughout the state space, and it all depends on which professor’s look to look. Those. between each point of Professor A and the point of Professor B is a contradiction.
He also understood that in order to resolve this contradiction, it is necessary to take into account the entire history as Professor A and Professor B put the experiment, what was happening around at that time. Many unnecessary and sometimes coincident data (after all, professors did almost the same thing). But only in them where there is a factor that affects.
Analyzing manually is tedious, so he set up a neural network and connected feedback links to it, which activated the history of the whole experiment. And if measuring the temperatures of the two sources after the forecast, we obtained data similar to those of Professor A, a reverse signal was sent about the context - the history of the experiment, by Professor A. And similarly with those close to Professor B. It was necessary to change the conditions of the experiment quite a lot did not receive similar to professorial patterns of data.
And here it is - and the network found the reason for the discrepancy, the network immediately began to work and it built a complete layout depending on another factor found, the pattern became more complex than that of the professors, but giving amazingly accurate data (though so far some third the professor will not get a contradiction) ... of course, we still need to choose this pattern from the network and write it analytically, but here it will be necessary to turn to the guy who sprawled out on the bench next to the gazebo ... but no, it turned out to be a dream, our dropout -geroy dozed off until the network properly spun, and he went to drink beer, thinking about the future of the universe.
1. Model of functional separation of consciousness and unconscious. Introduction
2. Model of the manifestation of consciousness or ANN without the effect of forgetting
3. The problem of "two or more teachers." First touches
4. Training with reinforcements on neural networks. Theory
The audience seemed to have the necessary knowledge, all the same, we have the color of society - programmers :) But alas ... the last survey showed that not all programmers are aware of AI issues. And boorish students who have run into these articles in the comments have not yet completed their studies.
')
Let's try to summarize and consider this extended article promised in the survey. And at the same time I was told that there are serious problems in the AI community. After a number of comments - yes, apparently there really is. Let's try to look at the trend.
Subjective opinion about the AI-community
Let's start with this. Many people will not like it, but try to soberly evaluate it yourself.
As in any science, AI has already acquired a full-fledged infrastructure :). And this is what it looks like:
1. At the foot of the AI pyramid, oracles sit - these are the people who loudly declare that we are engaged in "strong AI", "weak AI" - we are not interested. These freaks - at least the ones I met were all freaks. These are creationists of the 21st century, they adopt logic and try to build fictitious theories. No more no less. If you are a young man who is still not tied up - get out of this swamp, it sucks good smart heads. At first, do not even talk to such people - they feed on your energy, like religious sects.
2. Engineers are located on the steps to the AI pyramid. They do not stop to engage in "weak AI", often it is their bread. Even at the institute, they choose a very narrow direction - recognition of machine numbers, image processing, stock prediction, robot control, etc. etc. Often they have the title of doctor of science, and a little less often have the title of professors. If you are a young man who got to work with such engineers, or listen to their lecture - soak up the basics, otherwise you will certainly then slide to the foot. But having absorbed the basics - do not get carried away. Do not become a student who knows only the terminology - express your thoughts naturally. Try to do something yourself, never judge an algorithm, an approach only on the basis of what such an engineer said to you. It often turns out that he just read a lot of books and judges them (yes, I knew good AI professors who admitted this, which only did them more honor). If you believe him in what he does not directly deal with, then without checking the originals, in a couple of years you will transfer this knowledge from hearsay. As a result, you'll make a little more than a mistake. Of course, to check the originals and try to reproduce the experiments of the authors of the methods is time consuming. But otherwise, over time, you will have to believe a false impression that you understand something in the neighboring areas. And you will no longer become just a cheeky student who knows only terminology, but a person with authority, because you were engaged in a private (you will proudly call her a specific and concrete result) task, but at the same time you will distribute obviously inaccurate impressions that did not try experimentally.
3. On the steps in the shade, a little away from the heat, the neurophysiology engineers crouched, they got so tired of doing their work with patients, that to think about the other side, the AI engineers shout - no longer possible. They quietly take and do in their experiments, something that can be disassembled in the distant hum of the hassle of AI engineers. Something that they have, that there is no ... but as a rule they do not keep up with the thought of AI-engineers. But due to the fact that they really embody it in a human-machine interface, it always looks impressive.
4. But the question is - who is located in the gazebo on the top of the pyramid? It is not strange - it is empty, no input is free, just no one goes there. Sometimes only fleeting engineers in the heat of a fight fly in there, and also quickly leave. This is a place for AI theorists. Engineers often do not distinguish where the bottom, and where up - so they confuse theorist with the oracle. But this is how the brain of an AI engineer works; he, seeing theoretical generalizations, doesn’t see a specific task and doesn’t understand if it’s a bug. He does not see the forest for the trees, and seeing the sun, he takes it for the revelation of the oracle - and he thinks, ugh, why should I go up to the sun to do something not concrete. So he does not rise above the trees, fearing to be burned by the sun. And those who burned - he sees - groveling under his feet, on the outskirts of the pyramid. But of course, if he studied from the words of his contemporaries - he doesn’t know that those who created the concept of neural networks, genetic algorithms, fuzzy logic and much more ... still visited this gazebo at the top of the pyramid, just not very accepted speak.
Theorist's work is hard and not grateful.
When at last it happens to push through a crowd of AI engineers and an unenviable fate awaits an AI theoretician in the arbor. From above it becomes clearly visible that he does not have the right for simple solutions to simple problems. He can no longer afford to take a private, but previously favorite, specific task. He sees that in order not to be pushed out of the gazebo, he needs to cool down a little the heat of the crowd on the steps of the pyramid. And you need to splash water on everyone and not on any separate group.
But how can you convince that the splashed water is suitable for drinking by AI-engineers? After all, they only need gadgets to solve specific problems. Here the most difficult. It is necessary to degenerate the tasks of AI-engineers, only then they will be something common. And find a solution to this degenerate problem. Further, in fact, it is not lordly to adapt degenerate tasks to the real world from which AI engineers derive their tasks. They need to be given at least a tool, and already their problem is to use a hammer or an ax.
But it is becoming more and more difficult for the modern theorist, unpretentious degenerations have already been built - there are already ideas about classification problems, clustering problems, MDP, PoMDP, etc. We have to degenerate tasks so that they would increasingly remain the problems of the real world.
So we consider one such degenerate problem.
The necessary minimum for understanding is how to set the fitness function using an artificial neural network, and why?
At first I’ll have to take off my rose-colored glasses from those who wear them. Let's talk about the ability of the INS to predict and the speed of their learning. Everything is in accordance with Minsky - they forecast badly and slowly. And although I will be showing on the Rosenblatt perceptron, there is no fundamental difference for other networks.
This is for you AI engineers select a beautiful sample of tasks with good results on specific tasks. I'll show you the situation in the forehead. Who does not agree - we take the task described below and test it ourselves and provide specific results, with the code laid out - so that you can reproduce the results (I am tired of boltology - have entered).
Task.
We take a simple function c = a + b, where a and b are integers from 1 to 64, and c is an integer from 1 to 256 (we will need a slightly larger dimension later).
Then we will have 16 inputs and 256 outputs in the perceptron. Those. At the input we will submit 2 bytes representing in binary form the numbers a and b. And in order for the perceptron to work a little better, the output will be interpreted as what the output number is and the number. To get rid of the ambiguity of the way out - we work on the principle of "the winner takes all", i.e. We consider only the output that is most active.
The idea of the accuracy of the prediction of the neural network
If we teach the perceptron throughout the specified space of the inputs of the outputs, then it will unmistakably learn how to fold. To display this graphically, we draw a “red square” :) where, starting from the upper left corner along the X axis, put the number a, along the Y axis b, and the sum c at the intersection is displayed with a dot, and in the gradation of red. We get the following:
But when everything is known there is no place for prediction. So let's remove every second point:
Now take any graphics editor and complete the missing points. Can you restore the "red square". I think that only one person can do it - yes, yes to Malevich :)
But the perceptron will make a mistake just like any other person at almost every point, but will retain only the general tendency. Here is what he will do:
Thus, we see when mathematical accuracy is needed - predictions are fortune telling on the coffee grounds. But as an indication of some tendency is quite suitable. Therefore, the perceptrons work with image processing, since Our eye can be deceived, and detailed accuracy is not needed there. But in machines where the arithmetic precision is needed, the perceptron will not work. The same property and our intuition is with you - it can be said the use of our graphic memory in analytics. That's the whole psychology.
Fitness function
It's actually easy. The above function c = a + b we use may well reflect suitability for some process. (In economics, they like to talk about the utility function, they have everything there that is not useful, they don’t have to spend money :), or they’re also talking about the evaluation function - these are also fans of prices :) Well, we don’t ask for natural processes - we judge the suitability of suitable conditions for our actions). Well, let us heat the water in the kettle with two sources, separately from which are the temperature sensors of each source. We need to determine when the water boils.
in fact, we solved this problem above. Above, we trained the perceptron to perform the fitness function c = a + b. It remains only to beautifully display the temperature above 100 degrees, let's do it in the gradation of green that's what we get.
And we see that in general, where the perceptron has not abandoned the forecast (black dots), it is not wrong in principle in the trend. AI engineers can play around with parameters, types of neural networks, etc. and get almost no mistaken picture. In general, there will be many mistakes - but the border will be found. And thus completely fuzzy device will be able to give guidance to the action quite clearly. This is used by those who teach agents to run on the basis of a particular fitness function.
Now you just could not understand how to set the fitness function on an artificial neural network;)
Why is it convenient to set the fitness function using an artificial neural network?
This question is somewhat more complicated. It would seem, why use such a network that is inaccurate and slow in learning , when you can easily set the fitness function with a mathematical formula? All the same, that not to use intuition, and to use the bases of analytics received at school.
The problem with the mathematical formula is only one - it is tough and uncompromising.
But in practice, we began to measure the temperature and in previously unknown points - we began to receive some kind of anomaly. It is not known why, but they started. In the region of still quite small temperature sources - the temperature became sharply above 100 degrees. Here’s how it might look:
And now let's imagine each dimension is given to us very dearly, it takes several days and $ 10,000 - and understand what the anomaly is, and at least in what area we really want. Our formula is irrevocably obsolete and not suitable.
And here nothing is better and did not figure out how to use the neural network, although it can be bad, it can give a forecast - at least like this:
The problem of "two or more teachers" in the context of
It will be a little more difficult. Get ready to think :), I guess that the above described became clear even to the watchman in the market.
Suppose now we have two professors, one of them measured the heating of our water from two sources and received a layout (on the right), and the second layout (on the left):
Then one analyzed and built the theory that the temperature of the sources must be added to get the final temperature. And the second built another theory - it is enough to take only one of the temperatures of the source that is hotter.
And so they met, of course they quarreled, calling each other ignorant, deceivers, etc. and did not agree on anything, each went to his laboratory. Ordinary story is it? Then each of them created his own “School”. He began to teach students to his theories, and each boorish student of one school argued with the same boorish student of another school - after all, they had to shout over each other - so students win a place in the sun on the steps of our pyramid.
But it was clear to one dropout for a long time watching the struggle of these schools. But there is probably some factor that the professor does not take into account both school A and school professor B. He decides to tell them, but they closed in the study, and they are not allowed people on the street, afraid to let in some oracle. Therefore, our dropout has to wade through the students of professors, they of course each chuckled in his own way, twisted his finger at his temple, and one even strove to drop by the nose. Realizing that things will not be, having matured and pootbivy side of our hero retired to the gazebo at the top of the pyramid. There he realized that until he showed what the factor of the divergence between the two professors would be, he would not be listened to. And then he looked at the people who were seated and crowded on the steps. And I understood - not, since I had already climbed so high - why should I try on two professors, how many of them argue with each other and each couple has a particular problem - the uncertainty of the general factor of divergence of theories.
And he began to think about the problem in general categories, then to sprinkle all those sitting on the steps of the pyramid. Yes - he became infected with the spirit of the arbor, and I understood why many people do not want to go here. It really breaks a person.
And that's what he understood for now. Using the neural network as described above to identify where the anomaly is fails. Anomalies are everywhere throughout the state space, and it all depends on which professor’s look to look. Those. between each point of Professor A and the point of Professor B is a contradiction.
He also understood that in order to resolve this contradiction, it is necessary to take into account the entire history as Professor A and Professor B put the experiment, what was happening around at that time. Many unnecessary and sometimes coincident data (after all, professors did almost the same thing). But only in them where there is a factor that affects.
Analyzing manually is tedious, so he set up a neural network and connected feedback links to it, which activated the history of the whole experiment. And if measuring the temperatures of the two sources after the forecast, we obtained data similar to those of Professor A, a reverse signal was sent about the context - the history of the experiment, by Professor A. And similarly with those close to Professor B. It was necessary to change the conditions of the experiment quite a lot did not receive similar to professorial patterns of data.
And here it is - and the network found the reason for the discrepancy, the network immediately began to work and it built a complete layout depending on another factor found, the pattern became more complex than that of the professors, but giving amazingly accurate data (though so far some third the professor will not get a contradiction) ... of course, we still need to choose this pattern from the network and write it analytically, but here it will be necessary to turn to the guy who sprawled out on the bench next to the gazebo ... but no, it turned out to be a dream, our dropout -geroy dozed off until the network properly spun, and he went to drink beer, thinking about the future of the universe.
Source: https://habr.com/ru/post/148849/
All Articles