March 19, 2008 at 17:19Artificial Intelligence History, Part 2. Neural Network AI - Inevitable or Impossible?
When I was just thinking of writing this article, I only knew about neural networks that they kind of copy the thinking process of our brain. I did not suspect how wrong I was then.
At a time when cybernetics were just starting to play their toys, other more serious scientists were working on a more serious problem. Based on neurophysiological data on the structure of neurons, the cells of our brain, they tried to recreate their structure. It was a few years before the seminar where they first talked about AI.
These serious scientists were one hundred percent sure that the only thing that can think is our brain. Therefore, everything that should have this ability should reproduce the structure of our brain. A bold statement, especially when you consider that they had a rather vague idea about the processes of thinking, although no, they were not an idea, but simply a hypothesis that human thinking works through its own neural networks.
In other words, they, knowing the structure of a neuron, created their simplified copy of it and, knowing how neurons unite in a network, they tried to develop logic based on this knowledge, and then a computer. And only mathematicians work here. No one says that a person uses his neurons in the same way as a neurocomputer. It's just that engineers replaced the classic semiconductors with artificial neurons and are trying to build a new logic. No one is thinking about the psychological side of the process.
The principle of operation of neural networks is somewhat different from semiconductor circuits - the former do not use binary codes. But the secret weapon of neural networks is that neural networks are rumored to be able to learn. Let's try to figure this out.
So, we have a network (single-layer or multi-layer), consisting of simplified models of neurons. There is a group of input neurons and, accordingly, a group of output. When applying information to the input, the signals run through the network, amplifying and weakening, from neuron to neuron. They can go directly from entrance to exit, but they can also make loops, returning several neurons back. They may even randomly move throughout the network, but the point is that correctly processed signal is formed at the outputs.
If the coefficients of the quadratic equation are applied to the inputs, then at the output we must obtain its solutions. If at the inputs there is an image of handwritten text, then its recognized variant should be formed at the output.
At the beginning of work, when the neural network is not yet configured, there is no sense in it. In order for her to be useful in solving problems, besides her mysterious name, she needs to be trained. This process begins with the fact that the problem is solved using some external means. This is necessary in order to eventually get sets of questions and the right answers. Further, these questions are submitted to the input, and the answers to the exit.
Consider how this process goes for one neuron. The input signals come to its inputs, and the signal comes to the output, which is known in advance that it is correct. The neuron, in turn, forms its output signal, summing the input and applying a special function to the sum.
As a result, the intrinsic output signal of the neuron at the first stage differs from the external one. If a discrepancy is observed, then special algorithms are applied that change the level of the input signals (their weight) until both output signals are close.
But here you need to be careful, because if they are very close, the neural network will “relearn”. And then, if its task consists, for example, of text recognition, then the neural network will be able to recognize only the handwriting of a specific person, but not others. And if these two signals are very different, then the neural network will “underestimate”. That is, it will not be able to recognize anyone’s handwriting at all.
To prevent this from happening, the developer who is engaged in the training establishes a kind of tuning error. If the difference between the output signals does not exceed the value of this error, then the learning process can be considered complete.
Something like this ... In fact, everything is much more complicated, since the network has many neurons. The networks are different, and no one yet knows what it really should be, moreover, no one knows how many neurons there should be in it.
No matter how complicated and unimaginable all this seemed, the first neural network computers were built 50 years ago. And although they were not very powerful, they worked. The main differences between these computers from ordinary ones are self-learning (although I would prefer the term "self-tuning") and parallel computing.
It is because of this that they have such high hopes, since our ordinary processor computers do all the calculations sequentially. Parallel computing gives a new impetus to the development of search programs, cryptanalysis and other areas related to enumeration of a large amount of information.
Now neural networks are used for pattern recognition, text, speech. There are even those on the basis of which entire software systems are built that replace traders on the exchange. Expansion cards with neural networks are produced, which can then be inserted into a regular computer and work with them yourself with a set of the necessary software. They do their job well, although they are still in the process of improvement. But is it possible to create Artificial Intelligence on their basis?
Yes, the neural network is capable of self-study, but for this it is required to know all the answers in advance. The trial and error method will not work here. So this can only be called self-study with a big stretch. It can be configured to solve a specific problem. She will then deal with her, but with nothing else. To retrain her for new purposes, you need a person, and not just a person, but an expert in this new field. The neural network itself is not able to learn anything.
The trick of neurocybernetics was that they are trying to recreate the brain with its thinking processes. But the only thing they achieved was that they partially repeated the structure of only one type of its cells - neurons. At the same time, they screwed their algorithms to all this, which have nothing to do with our thinking.
As a result, we get the same cybernetics of the "black box", with input, output and hell than inside. The process of our thinking has nothing to do with the work of a neural network, since neural networks do not use analysis, synthesis, comparison, deduction. Also, they do not take into account such a factor as the human psyche. When neurons are recreated, their properties associated with all this are omitted, so the process is realized only at the surface level.
The creators of neural networks, who are still trying, in addition to logic, to screw some kind of “programming language” based on human psychology, faced a serious problem. They are trying to find a general theory of psychology that could be easily formalized. But the fact is that there are dozens of approaches here, and what is even more interesting is that not one of them refers to physiology, that is, to the fact that our neural networks are the engine of our psyche. But what if there is no such connection?
Theoretical psychology is already an independent system with its own elements. It is not attached to physiology in any way, therefore this “programming language", this formalization of our psychology can be implemented both on neural networks and on another platform, or generally described mathematically. Therefore, neural networks are not so irreplaceable, since you can do without them.
It is time to return to the question that I asked at the end of the first part, namely the question of whether the existing definition of Artificial Intelligence is so good. It seems like all the components have already been created, but something is still missing for this Frankenstein. There must be something that breathes life into him.
The main thing here is to understand why we need Artificial Intelligence. Solve logic problems? An ordinary computer can already serve for this. Recognize images or speech? Such technologies also already exist. Maybe there is some other task that is quite complicated? I think that for her there is some technology or development. Any of our problems can be solved separately and without involving Artificial Intelligence. Then why do we need it?
I suspect not for any specific task. We will not beat around the bush and deceive ourselves. Let's say directly that we want Artificial Intelligence to be as close to human as possible.
So that he was equally illogical, possessed intuition, the ability to give rise to ideas, make decisions, feel, empathize, so that when communicating with him a feeling is created that you are communicating with a person. So that he has his own view of the world, so that he can argue and agree or disagree with his opponent. That he was attentive, had the ability to be friends, keep secrets, lie, respect and dislike, so that he had a sense of humor. So that he can love.
It seems to me that this definition is closer to the point.
UPD: Judging by the comments, it became clear that the name does not reflect the essence, so it had to be changed.
Table of Contents:
The History of Artificial Intelligence, Part 1. Painting without an artist.
Artificial Intelligence History, Part 2. Neural Network AI - Inevitable or Impossible?
Making Artificial Intelligence
At a time when cybernetics were just starting to play their toys, other more serious scientists were working on a more serious problem. Based on neurophysiological data on the structure of neurons, the cells of our brain, they tried to recreate their structure. It was a few years before the seminar where they first talked about AI.
These serious scientists were one hundred percent sure that the only thing that can think is our brain. Therefore, everything that should have this ability should reproduce the structure of our brain. A bold statement, especially when you consider that they had a rather vague idea about the processes of thinking, although no, they were not an idea, but simply a hypothesis that human thinking works through its own neural networks.
In other words, they, knowing the structure of a neuron, created their simplified copy of it and, knowing how neurons unite in a network, they tried to develop logic based on this knowledge, and then a computer. And only mathematicians work here. No one says that a person uses his neurons in the same way as a neurocomputer. It's just that engineers replaced the classic semiconductors with artificial neurons and are trying to build a new logic. No one is thinking about the psychological side of the process.
The principle of operation of neural networks is somewhat different from semiconductor circuits - the former do not use binary codes. But the secret weapon of neural networks is that neural networks are rumored to be able to learn. Let's try to figure this out.
So, we have a network (single-layer or multi-layer), consisting of simplified models of neurons. There is a group of input neurons and, accordingly, a group of output. When applying information to the input, the signals run through the network, amplifying and weakening, from neuron to neuron. They can go directly from entrance to exit, but they can also make loops, returning several neurons back. They may even randomly move throughout the network, but the point is that correctly processed signal is formed at the outputs.
If the coefficients of the quadratic equation are applied to the inputs, then at the output we must obtain its solutions. If at the inputs there is an image of handwritten text, then its recognized variant should be formed at the output.
At the beginning of work, when the neural network is not yet configured, there is no sense in it. In order for her to be useful in solving problems, besides her mysterious name, she needs to be trained. This process begins with the fact that the problem is solved using some external means. This is necessary in order to eventually get sets of questions and the right answers. Further, these questions are submitted to the input, and the answers to the exit.
Consider how this process goes for one neuron. The input signals come to its inputs, and the signal comes to the output, which is known in advance that it is correct. The neuron, in turn, forms its output signal, summing the input and applying a special function to the sum.
As a result, the intrinsic output signal of the neuron at the first stage differs from the external one. If a discrepancy is observed, then special algorithms are applied that change the level of the input signals (their weight) until both output signals are close.
But here you need to be careful, because if they are very close, the neural network will “relearn”. And then, if its task consists, for example, of text recognition, then the neural network will be able to recognize only the handwriting of a specific person, but not others. And if these two signals are very different, then the neural network will “underestimate”. That is, it will not be able to recognize anyone’s handwriting at all.
To prevent this from happening, the developer who is engaged in the training establishes a kind of tuning error. If the difference between the output signals does not exceed the value of this error, then the learning process can be considered complete.
Something like this ... In fact, everything is much more complicated, since the network has many neurons. The networks are different, and no one yet knows what it really should be, moreover, no one knows how many neurons there should be in it.
No matter how complicated and unimaginable all this seemed, the first neural network computers were built 50 years ago. And although they were not very powerful, they worked. The main differences between these computers from ordinary ones are self-learning (although I would prefer the term "self-tuning") and parallel computing.
It is because of this that they have such high hopes, since our ordinary processor computers do all the calculations sequentially. Parallel computing gives a new impetus to the development of search programs, cryptanalysis and other areas related to enumeration of a large amount of information.
Now neural networks are used for pattern recognition, text, speech. There are even those on the basis of which entire software systems are built that replace traders on the exchange. Expansion cards with neural networks are produced, which can then be inserted into a regular computer and work with them yourself with a set of the necessary software. They do their job well, although they are still in the process of improvement. But is it possible to create Artificial Intelligence on their basis?
Yes, the neural network is capable of self-study, but for this it is required to know all the answers in advance. The trial and error method will not work here. So this can only be called self-study with a big stretch. It can be configured to solve a specific problem. She will then deal with her, but with nothing else. To retrain her for new purposes, you need a person, and not just a person, but an expert in this new field. The neural network itself is not able to learn anything.
The trick of neurocybernetics was that they are trying to recreate the brain with its thinking processes. But the only thing they achieved was that they partially repeated the structure of only one type of its cells - neurons. At the same time, they screwed their algorithms to all this, which have nothing to do with our thinking.
As a result, we get the same cybernetics of the "black box", with input, output and hell than inside. The process of our thinking has nothing to do with the work of a neural network, since neural networks do not use analysis, synthesis, comparison, deduction. Also, they do not take into account such a factor as the human psyche. When neurons are recreated, their properties associated with all this are omitted, so the process is realized only at the surface level.
The creators of neural networks, who are still trying, in addition to logic, to screw some kind of “programming language” based on human psychology, faced a serious problem. They are trying to find a general theory of psychology that could be easily formalized. But the fact is that there are dozens of approaches here, and what is even more interesting is that not one of them refers to physiology, that is, to the fact that our neural networks are the engine of our psyche. But what if there is no such connection?
Theoretical psychology is already an independent system with its own elements. It is not attached to physiology in any way, therefore this “programming language", this formalization of our psychology can be implemented both on neural networks and on another platform, or generally described mathematically. Therefore, neural networks are not so irreplaceable, since you can do without them.
It is time to return to the question that I asked at the end of the first part, namely the question of whether the existing definition of Artificial Intelligence is so good. It seems like all the components have already been created, but something is still missing for this Frankenstein. There must be something that breathes life into him.
The main thing here is to understand why we need Artificial Intelligence. Solve logic problems? An ordinary computer can already serve for this. Recognize images or speech? Such technologies also already exist. Maybe there is some other task that is quite complicated? I think that for her there is some technology or development. Any of our problems can be solved separately and without involving Artificial Intelligence. Then why do we need it?
I suspect not for any specific task. We will not beat around the bush and deceive ourselves. Let's say directly that we want Artificial Intelligence to be as close to human as possible.
So that he was equally illogical, possessed intuition, the ability to give rise to ideas, make decisions, feel, empathize, so that when communicating with him a feeling is created that you are communicating with a person. So that he has his own view of the world, so that he can argue and agree or disagree with his opponent. That he was attentive, had the ability to be friends, keep secrets, lie, respect and dislike, so that he had a sense of humor. So that he can love.
It seems to me that this definition is closer to the point.
UPD: Judging by the comments, it became clear that the name does not reflect the essence, so it had to be changed.
Table of Contents:
The History of Artificial Intelligence, Part 1. Painting without an artist.
Artificial Intelligence History, Part 2. Neural Network AI - Inevitable or Impossible?
Making Artificial Intelligence