Esperanto for robots and Smith agents
William Shakespeare Dictionary is 12,000 words. The dictionary of the Negro from the cannibalistic tribe "Mumbo-Yumbo" is 300 words. Ellochka Shchukin from the book “12 chairs” easily and freely managed thirty.
The man easily understands Ellochku. Alas, the computer would rather understand Shakespeare than Ellochku. We have enough to understand the context and intonation - we will think the rest, - and the computer needs clarity in every term. Preferably without synonyms. The more images hidden behind each word, the harder it is to understand. The context-dependence of human words is only one of the difficulties of learning artificial intelligence in natural language. If the edible onion is distinguished from the shooting onion, if the situation is completely described, then denial of meaning with irony or rhetorical questions is recognized only by intonation. Additional complexity is created by the syntax and word order in the sentence.
Igor Mordach, an employee of the American laboratory of artificial intelligence OpenAI, however, decided that Ellochkin language is far from a dead-end path, but a perfectly suitable example on which bots will build a language for their communication, and then they will understand us. Only 30 words for computer Ellochka he regretted. In his experiment, a computer program created a language from scratch.
')
The accuracy of natural language recognition is now among the leaders of cognitive systems (IBM Watson, Google, ABBYY, Microsoft, Nanosemantics) allows you to generally understand the meaning and answer the written question with a predetermined subject knowledge base, but the conversation even with 90% accuracy of phrase recognition actually very tiresome. Conversation, full of interjections and often taking place in a large company with the intersection of dialogues, puts the bot on the ability to keep the conversation up to the level of a three-year-old child.
Having the ability of programs to accurately respond to the meaning of phrases would solve many problems of the interface, and thus connect bots (artificial intelligence agents) to any everyday human task. There is a dialogue with cars even now - TV, for example, communicates in the language of the console. But he understands very few teams. And they understand them only after programming. From this simple example it is clear that for quick and flexible communication with multi-purpose programs and devices (at least robots), it is necessary to remove intermediaries from the dialogue. Among which, alas, and programmers.
Computer linguistic optimists pin their hopes on neural networks. In their understanding, the problem is only in the need to process a large number of examples. Opponents consider the task intractable. Not so much because of the scale of training, but because of the equivalence of this task of the task of creating consciousness in machines. Examples of counting dolphins or communicating monkeys each side treats itself to its advantage - some as a prospect, others as a dead end. For those interested in the details on Habré and GT, there are many posts about natural language. For example, "Introduction to the recognition of natural language . "
The subject of this post is the third approach, taken as a basis by Igor Mordach - let the machines themselves first learn to communicate in the language they create. The process of developing your own language will allow software agents of artificial intelligence to better understand the algorithm for adjusting terms to new meanings, the rules of phrase formation and use this knowledge to talk with a person. This is exactly how Mordach set the task, who previously worked as a specialist in the creation of moving movie heroes. With this background in mind, Igor decided that movement training contains algorithms for collecting information and its partial simultaneous use, which can be used in any training.
To teach bots the language of bots (software agents), they were placed in the conditional universe of the “white square”, they were given goals, energy of movement and the ability to process the experience of relating themselves to the goal. Initially, bots had a minimal set of action commands, personal experience of success, and also provided a way to exchange information - through close "tactile contact" or through a remote prototype of "hearing and vision" (here it is more appropriate to use the term information flow direction). The purpose of the bots in the conditional universe of the “white square” was to reach a certain point on their own, communicate their goal to another agent, exchange goals, push another agent to the goal if there is no information exchange between them. Targets and bots differed in color or size.
For Agent Smith, when created by Mordachem, the matrix would look something like this:
In the OpenAI experiment, the fact and speed of achieving goals was the measured parameter of success of trial and error. They could be inaction or movement of a certain point (direct contact with the target or contact through other agents) in the virtual world of the white square. The faster the goal was achieved, the more useful the list of terms used was recognized in the current dialogue of goals. Goals and success were both individual and collective.
In the process of performing tasks, bots have enriched the language reserve with new concepts that conditionally correspond to human concepts, from movement terms to concepts: above, below, to the right, to the left. Mordach expects that by establishing the rules of word formation, bots can more easily “understand” the same rules in human languages. Including learning the accuracy of using comparative and indefinite terms. AI poorly understands abstract and generalized terms. More precisely, he does not understand anything at all, but with some phrases he is ready to work, but for others it is necessary to get clarifying information. In the expression "the car goes fast" at least the word "fast" - needs to be measured. Even in terms of traffic rules, “fast” is still an array of values ​​that fit into some interval. More precisely, several intervals - for the city, highways and rural areas.
Neural networks, of course, are able to work with indefinite terms, but their training usually does not start from scratch, contains a number of starting processing rules, and also has no dependence of the accumulation of signaling experience on the type of channel through which information is transmitted. These are the main novelties of Mordach's experiments.
Comparison of remote communications and tactile communication of bots:
The absence of body and boundaries makes all these terms conditional, of course.
A remarkable point is that the success of each bot was calculated as a share of the success of all bots, encouraging cooperation in achieving the goal. That is, the quality of vocabulary was assessed by the totality of the language used. It turns out that someone had to necessarily collect and report the secondary circumstances of achieving the goal, which, in the case of individual fulfillment of the goals by each carrier, would prevent them from reaching them. Statistical rule 20/80 - receiving 20% ​​of information provides 80% of the goal achievement - works at the level of an individual subject, but in a team someone always has to give 100% of the collected information to an additional 1% of the goal achievement.
All cases of "verbal" and "non-verbal" contact communication bots in the study of Mordach shown in the video:
The main idea of ​​Mordach is to make sure that by comprehending the “surrounding world”, the bots create the language necessary for this comprehension. In his concept, language is the result of action. Other scientists pay attention to the fact that a developed language and the ability to communicate on abstract topics will allow machines to create their own ideas, to describe and create new entities that were not previously existing. Not far from the transformation of our world.
The ability to understand natural language is considered to be closely related to the presence of consciousness in the subject of communication. Will the ability to understand the meaning of what was said to the emergence of consciousness in cars? Perhaps only a person with a human-like creation can fully understand the natural language, and creating a language of bots in their own environment will lead to intelligent bots in one or another version of the Smith agent.
Setting bots priorities in the course of training is contrary to the engineering principle of reliability and predictability of the machines. The unpredictability of priorities, the ability to fantasize, self-study, the secrecy of communication between bots for humans - have already made themselves felt in previous OpenAI experiments . Without a mandatory final goal, the programs came up with unexpected priorities for themselves in the course of their studies. In the experiment on virtual boat races, the execution of related tasks became more profitable for bots than one place and generally passing the race to the end. Fulfillment of passing tasks brought 20% more points.
Familiar, is not it? Robert Sheckley described this 65 years ago in the famous story “The Warden of the Bird”, which went from preventing murders to blocking any activity on Earth, ignoring other priorities and tasks for the main purpose:
An excursion to create their own language for bots can be the creation of their own virtual universe. At least in the history of mankind such precedents exist. I quote Tolkien:
The fact that the machines will start to think before we understand this, fear in OpenAI. If bots have their own language, then people will miss this moment almost certainly. Therefore, Mordach has the task of not only creating the Boto language, but also creating a translator for the human. First, of course, in English.
Responding to such fears, Igor cites a well-known thought experiment in the field of the philosophy of consciousness and the philosophy of artificial intelligence called the Chinese Room , published by John Pearl in 1980, with a counter argument. The essence of Searl’s statement is that any full-fledged communication is theoretically quite possible without consciousness, but using sufficiently detailed rules for interpreting the interlocutor’s questions for answers. A more general conclusion by Searle suggests that any manipulation of syntactic constructions cannot lead to understanding.
In such a context, its own language and its own virtual universe does not at all mean the presence of consciousness among its agents and objects. Like in computer games. Or, as some believe, like a person whose consciousness is only a semantic gateway between a clear “artificial” language inside the brain and the natural noise of signals from the outside world.
The man easily understands Ellochku. Alas, the computer would rather understand Shakespeare than Ellochku. We have enough to understand the context and intonation - we will think the rest, - and the computer needs clarity in every term. Preferably without synonyms. The more images hidden behind each word, the harder it is to understand. The context-dependence of human words is only one of the difficulties of learning artificial intelligence in natural language. If the edible onion is distinguished from the shooting onion, if the situation is completely described, then denial of meaning with irony or rhetorical questions is recognized only by intonation. Additional complexity is created by the syntax and word order in the sentence.
Igor Mordach, an employee of the American laboratory of artificial intelligence OpenAI, however, decided that Ellochkin language is far from a dead-end path, but a perfectly suitable example on which bots will build a language for their communication, and then they will understand us. Only 30 words for computer Ellochka he regretted. In his experiment, a computer program created a language from scratch.
')
In search of a universal language of communication with cars
The accuracy of natural language recognition is now among the leaders of cognitive systems (IBM Watson, Google, ABBYY, Microsoft, Nanosemantics) allows you to generally understand the meaning and answer the written question with a predetermined subject knowledge base, but the conversation even with 90% accuracy of phrase recognition actually very tiresome. Conversation, full of interjections and often taking place in a large company with the intersection of dialogues, puts the bot on the ability to keep the conversation up to the level of a three-year-old child.
Having the ability of programs to accurately respond to the meaning of phrases would solve many problems of the interface, and thus connect bots (artificial intelligence agents) to any everyday human task. There is a dialogue with cars even now - TV, for example, communicates in the language of the console. But he understands very few teams. And they understand them only after programming. From this simple example it is clear that for quick and flexible communication with multi-purpose programs and devices (at least robots), it is necessary to remove intermediaries from the dialogue. Among which, alas, and programmers.
Computer linguistic optimists pin their hopes on neural networks. In their understanding, the problem is only in the need to process a large number of examples. Opponents consider the task intractable. Not so much because of the scale of training, but because of the equivalence of this task of the task of creating consciousness in machines. Examples of counting dolphins or communicating monkeys each side treats itself to its advantage - some as a prospect, others as a dead end. For those interested in the details on Habré and GT, there are many posts about natural language. For example, "Introduction to the recognition of natural language . "
The subject of this post is the third approach, taken as a basis by Igor Mordach - let the machines themselves first learn to communicate in the language they create. The process of developing your own language will allow software agents of artificial intelligence to better understand the algorithm for adjusting terms to new meanings, the rules of phrase formation and use this knowledge to talk with a person. This is exactly how Mordach set the task, who previously worked as a specialist in the creation of moving movie heroes. With this background in mind, Igor decided that movement training contains algorithms for collecting information and its partial simultaneous use, which can be used in any training.
To teach bots the language of bots (software agents), they were placed in the conditional universe of the “white square”, they were given goals, energy of movement and the ability to process the experience of relating themselves to the goal. Initially, bots had a minimal set of action commands, personal experience of success, and also provided a way to exchange information - through close "tactile contact" or through a remote prototype of "hearing and vision" (here it is more appropriate to use the term information flow direction). The purpose of the bots in the conditional universe of the “white square” was to reach a certain point on their own, communicate their goal to another agent, exchange goals, push another agent to the goal if there is no information exchange between them. Targets and bots differed in color or size.
For Agent Smith, when created by Mordachem, the matrix would look something like this:
In the OpenAI experiment, the fact and speed of achieving goals was the measured parameter of success of trial and error. They could be inaction or movement of a certain point (direct contact with the target or contact through other agents) in the virtual world of the white square. The faster the goal was achieved, the more useful the list of terms used was recognized in the current dialogue of goals. Goals and success were both individual and collective.
In the process of performing tasks, bots have enriched the language reserve with new concepts that conditionally correspond to human concepts, from movement terms to concepts: above, below, to the right, to the left. Mordach expects that by establishing the rules of word formation, bots can more easily “understand” the same rules in human languages. Including learning the accuracy of using comparative and indefinite terms. AI poorly understands abstract and generalized terms. More precisely, he does not understand anything at all, but with some phrases he is ready to work, but for others it is necessary to get clarifying information. In the expression "the car goes fast" at least the word "fast" - needs to be measured. Even in terms of traffic rules, “fast” is still an array of values ​​that fit into some interval. More precisely, several intervals - for the city, highways and rural areas.
Neural networks, of course, are able to work with indefinite terms, but their training usually does not start from scratch, contains a number of starting processing rules, and also has no dependence of the accumulation of signaling experience on the type of channel through which information is transmitted. These are the main novelties of Mordach's experiments.
Comparison of remote communications and tactile communication of bots:
The absence of body and boundaries makes all these terms conditional, of course.
A remarkable point is that the success of each bot was calculated as a share of the success of all bots, encouraging cooperation in achieving the goal. That is, the quality of vocabulary was assessed by the totality of the language used. It turns out that someone had to necessarily collect and report the secondary circumstances of achieving the goal, which, in the case of individual fulfillment of the goals by each carrier, would prevent them from reaching them. Statistical rule 20/80 - receiving 20% ​​of information provides 80% of the goal achievement - works at the level of an individual subject, but in a team someone always has to give 100% of the collected information to an additional 1% of the goal achievement.
All cases of "verbal" and "non-verbal" contact communication bots in the study of Mordach shown in the video:
Hand-made and invisible Matrix as a result of Boto communication
The main idea of ​​Mordach is to make sure that by comprehending the “surrounding world”, the bots create the language necessary for this comprehension. In his concept, language is the result of action. Other scientists pay attention to the fact that a developed language and the ability to communicate on abstract topics will allow machines to create their own ideas, to describe and create new entities that were not previously existing. Not far from the transformation of our world.
The ability to understand natural language is considered to be closely related to the presence of consciousness in the subject of communication. Will the ability to understand the meaning of what was said to the emergence of consciousness in cars? Perhaps only a person with a human-like creation can fully understand the natural language, and creating a language of bots in their own environment will lead to intelligent bots in one or another version of the Smith agent.
Setting bots priorities in the course of training is contrary to the engineering principle of reliability and predictability of the machines. The unpredictability of priorities, the ability to fantasize, self-study, the secrecy of communication between bots for humans - have already made themselves felt in previous OpenAI experiments . Without a mandatory final goal, the programs came up with unexpected priorities for themselves in the course of their studies. In the experiment on virtual boat races, the execution of related tasks became more profitable for bots than one place and generally passing the race to the end. Fulfillment of passing tasks brought 20% more points.
Familiar, is not it? Robert Sheckley described this 65 years ago in the famous story “The Warden of the Bird”, which went from preventing murders to blocking any activity on Earth, ignoring other priorities and tasks for the main purpose:
Jackson yawned and stopped the car at the curb. He did not notice a sparkling point in the sky. He had no reason to beware. After all, by all human notions, he did not intend to kill.
He extended his hand, wanted to turn off the ignition ... And something threw him against the wall of the cabin.
“Iron fool,” he said to a mechanical bird hanging above him. The car is not alive. I don't want to kill him at all.
But the bird watchdog knew one thing: some actions stop the activity of the organism. The car is definitely an active organism. After all, it is made of metal, just like the guard bird itself, isn't it? And while moving ...
An excursion to create their own language for bots can be the creation of their own virtual universe. At least in the history of mankind such precedents exist. I quote Tolkien:
The invention of languages ​​is the basis of my works. "Stories" were written more to create a world for these languages, and not vice versa. For me, the word first appears, and then - the story associated with it. I would prefer to write "elvish." But, of course, such a book as “The Lord of the Rings” has undergone serious editing, and I left only as many “languages” there as the reader could digest (although now I find out that many would like more). [...] In any case, for me this is largely an essay on “linguistic aesthetics,” as I sometimes answer to people who ask me what I wrote my book.
The fact that the machines will start to think before we understand this, fear in OpenAI. If bots have their own language, then people will miss this moment almost certainly. Therefore, Mordach has the task of not only creating the Boto language, but also creating a translator for the human. First, of course, in English.
Responding to such fears, Igor cites a well-known thought experiment in the field of the philosophy of consciousness and the philosophy of artificial intelligence called the Chinese Room , published by John Pearl in 1980, with a counter argument. The essence of Searl’s statement is that any full-fledged communication is theoretically quite possible without consciousness, but using sufficiently detailed rules for interpreting the interlocutor’s questions for answers. A more general conclusion by Searle suggests that any manipulation of syntactic constructions cannot lead to understanding.
In such a context, its own language and its own virtual universe does not at all mean the presence of consciousness among its agents and objects. Like in computer games. Or, as some believe, like a person whose consciousness is only a semantic gateway between a clear “artificial” language inside the brain and the natural noise of signals from the outside world.
Source: https://habr.com/ru/post/402407/
All Articles