Neurointelligence or myth?
The era of neural network artificial intelligence is in full swing! The real breakthrough in recent years - neural networks recognize not only speech, pictures with animals, complex scenes, but even describe them with words! (NeuralTalk). On neural networks, bots are made that are still bad, but they are already responding. Neural networks compose music and write poems. According to the scenario, neural networks are even going to make a movie! Soon, journalists predict, very soon they will drive cars, banks, corporations, countries. "And everyone will fly on the stage coach" - that was the dream at the beginning of the last century. So now the creators dream of a neural network future. They say that they will soon overtake a man in all. Neural networks have already beaten a man of chess, go, in Jaopardy (Own game). And so they call their creations no less than Artificial Intelligence. I recognized the horse in the picture - artificial intelligence. Why intelligence? Because before, only man could do it. It is strange why, then, the calculator was not named artificial intelligence. Is it then so called neural networks?
Of course, a neural network is orders of magnitude more complicated than a calculator. But, if you look at the result of the work of a neural network without anticipating the intellect, it is just a classifier! Moreover, the regressive nature - the set of inputs it reduces to the choice of several options (except for associative networks and the Boltzmann machine). Relay, only with more complex input. It seemed to me that intelligence is at least the ability to reason. And while no network can build at least the simplest syllogism, to call it intelligence, to put it mildly, a bit early. I'm not talking about more complex tasks. For example,
A bee has fewer neurons in the head, but it can do something that no modern neural network is capable of yet - fly at high speed among branches and trees. So far Google hardly sweats up over 10 years of development and drives cars along road markings. And the bees are an insect, the very beginning of the evolution of a neural network. We have not even reached this initial level, but already call networks the creation, which marked the crown of the evolution of the neural network - intelligence.
All, and scientists themselves, make the same epistemological error. They say that the network has intelligence because it recognized the horse in the picture! Did she recognize the horse? She simply chose to exit X, which the researcher called "horse." Yes, she made a generalization (classifier), reducing the version of the horse to the “horse” exit. But the neural network has no idea what it has chosen. She has no idea about the "horse" (semantics), the "meaning" of the output "horse" in our head, not in the network. In order for it to be in the neural network itself, it must choose not an input, not a word according to the recognized picture, but all contextually related words, that is, the whole concept of "horse." Then she may understand that the horse and the nebula have a general meaning only in the sense of a cloud of cosmic dust. For a neural network, the selected output is a semantic “point” that has no content.
')
The weakest achievements of neural networks are in the NLP, where understanding of a natural language is required. Because we still do not seem to understand very well what it means to “understand”. And this is the most important aspect that we would like to implement in the car. And which could then rightly be called intelligence. Understanding precisely in the sense of the semantics of the "horse." At the same time, it seems that we are approaching the threshold of understanding what we are doing - the networks create, train, but they cannot understand how they achieve results by the recognition of the creators themselves. Although we have created only the first brick in the building of Intellect. Can we then reach a more serious level of AI, which is more complicated than the current one by an order of magnitude?
Total
As a result, I will quote from an article that has just appeared on Habré about the “successes” of neural networks in the language (semantic) imitation of habrahabr.ru/company/payonline/blog/307666
“There is only one problem that quickly becomes apparent as you monitor other system responses. When Lee asked, “How many legs does a cat have?”, The system responded: “Four, I guess.” After that, he made another attempt: "How many legs does a centipede have?" The answer was curious: "Eight." In essence, Lee's program has no idea what she is talking about. She understands that certain combinations of characters can occur together, but she has no idea about the existence of the real world. She doesn't know what the centipede really looks like, or how it moves. That is, we still have only an illusion of intellect , deprived of the very common sense that we humans perceive as something taken for granted. Such instability of the results is quite common for depth learning systems. Google’s signature program makes weird mistakes. Looking at a road sign, for example, she might call it a food-filled refrigerator. ”
From the comments, I note ZhenyaZero, which very precisely expressed the specific difference between modern neural networks and how we, humans, recognize the pictures. “Nevertheless, you will undoubtedly distinguish a horse with a fifth foot from a tiger with a fifth foot. And when describing a picture, most people will say “a horse with a fifth leg,” and not “I think this is a piece of cake.” And the results of a neural network on unusual and borderline variants are really poorly predictable and often look completely inadequate.
And also a few facts.
1) To see so many different "horses", not a single person will have enough of a lifetime (s).
2) It is often enough for a child to show once a picture of a new animal and explain how it differs so that he can recognize such animals the next time and in any poses.
3) We mean "horse" even when we see just a bridle.
The conclusions of the article are quite obvious, but it is more important to understand that why such differences exist, and what to do next. Your suggestions.
Of course, a neural network is orders of magnitude more complicated than a calculator. But, if you look at the result of the work of a neural network without anticipating the intellect, it is just a classifier! Moreover, the regressive nature - the set of inputs it reduces to the choice of several options (except for associative networks and the Boltzmann machine). Relay, only with more complex input. It seemed to me that intelligence is at least the ability to reason. And while no network can build at least the simplest syllogism, to call it intelligence, to put it mildly, a bit early. I'm not talking about more complex tasks. For example,
A bee has fewer neurons in the head, but it can do something that no modern neural network is capable of yet - fly at high speed among branches and trees. So far Google hardly sweats up over 10 years of development and drives cars along road markings. And the bees are an insect, the very beginning of the evolution of a neural network. We have not even reached this initial level, but already call networks the creation, which marked the crown of the evolution of the neural network - intelligence.
All, and scientists themselves, make the same epistemological error. They say that the network has intelligence because it recognized the horse in the picture! Did she recognize the horse? She simply chose to exit X, which the researcher called "horse." Yes, she made a generalization (classifier), reducing the version of the horse to the “horse” exit. But the neural network has no idea what it has chosen. She has no idea about the "horse" (semantics), the "meaning" of the output "horse" in our head, not in the network. In order for it to be in the neural network itself, it must choose not an input, not a word according to the recognized picture, but all contextually related words, that is, the whole concept of "horse." Then she may understand that the horse and the nebula have a general meaning only in the sense of a cloud of cosmic dust. For a neural network, the selected output is a semantic “point” that has no content.
')
The weakest achievements of neural networks are in the NLP, where understanding of a natural language is required. Because we still do not seem to understand very well what it means to “understand”. And this is the most important aspect that we would like to implement in the car. And which could then rightly be called intelligence. Understanding precisely in the sense of the semantics of the "horse." At the same time, it seems that we are approaching the threshold of understanding what we are doing - the networks create, train, but they cannot understand how they achieve results by the recognition of the creators themselves. Although we have created only the first brick in the building of Intellect. Can we then reach a more serious level of AI, which is more complicated than the current one by an order of magnitude?
Total
As a result, I will quote from an article that has just appeared on Habré about the “successes” of neural networks in the language (semantic) imitation of habrahabr.ru/company/payonline/blog/307666
“There is only one problem that quickly becomes apparent as you monitor other system responses. When Lee asked, “How many legs does a cat have?”, The system responded: “Four, I guess.” After that, he made another attempt: "How many legs does a centipede have?" The answer was curious: "Eight." In essence, Lee's program has no idea what she is talking about. She understands that certain combinations of characters can occur together, but she has no idea about the existence of the real world. She doesn't know what the centipede really looks like, or how it moves. That is, we still have only an illusion of intellect , deprived of the very common sense that we humans perceive as something taken for granted. Such instability of the results is quite common for depth learning systems. Google’s signature program makes weird mistakes. Looking at a road sign, for example, she might call it a food-filled refrigerator. ”
From the comments, I note ZhenyaZero, which very precisely expressed the specific difference between modern neural networks and how we, humans, recognize the pictures. “Nevertheless, you will undoubtedly distinguish a horse with a fifth foot from a tiger with a fifth foot. And when describing a picture, most people will say “a horse with a fifth leg,” and not “I think this is a piece of cake.” And the results of a neural network on unusual and borderline variants are really poorly predictable and often look completely inadequate.
And also a few facts.
1) To see so many different "horses", not a single person will have enough of a lifetime (s).
2) It is often enough for a child to show once a picture of a new animal and explain how it differs so that he can recognize such animals the next time and in any poses.
3) We mean "horse" even when we see just a bridle.
The conclusions of the article are quite obvious, but it is more important to understand that why such differences exist, and what to do next. Your suggestions.
Source: https://habr.com/ru/post/396743/
All Articles