Why this winter we again invite everyone to play computer games using artificial intelligence
Habr, MIPT welcomes you! As true techies, we immediately get down to business and invite all who are interested to take part in the new hackathon DeepHack, which will be held at Fiztekh from February 6 to 12. The qualifying stage has already begun and will last until January 22. This is all we need ... If you know firsthand what DQN, deep RL and DeepHack are, immediately register for the next scientific school-hackathon - rl.deephack.me .
And if you are not completely in the subject line and, for example, it is not clear to you why computer games, how they relate to the management of data centers and what will actually be in February, then immediately go under the cat - there is a maximum immersion in the life of an artificial intelligence from antiquity to the present day. Well, you do not think that all this was invented only in the XXI century?
Cheat sheet to the reader: DeepHack is a format that combines a classic hackathon with pizza, Coca-Cola and round-the-clock coding and a scientific school with lectures from leading experts on the topic. What was told on past hackathons and what results turned out can be found on game.deephack.me and qa.deephack.me .
')
Delve into history. As you know, the first attempts to create artificial intelligence began a long time ago. The answer to the question, how long ago, strongly depends on what we are putting into the concept of “artificial intelligence”. We will assume that artificial intelligence is some kind of machine capable of making decisions using logical calculations. Such a definition will allow us to miss all (certainly very interesting) attempts to create artificial intelligence in Antiquity, Antiquity, the Middle Ages and the New Time, immediately skipping to the 19th century to Charles Babbage and Ada Lovelace. They are widely known as the creator of the computer and the first programmer, respectively. But let's not linger here and go even further - to the beginning of the 20th century, when David Gilbert, who became Henri Poincaré after the death as the universally recognized mathematician No. 1, put to the world mathematical community the question: “can all mathematical reasoning be formalized?” other well-known scientists gave this question: Kurt Godel with his incompleteness theorem (1931), Alan Turing, who proposed the car named after himself (1936) and Alonzo Church, who invented the λ-calculus (series of works in the 1930s). Gödel's work demonstrated that mathematical logic has its limits. But the results of Turing-Church showed that everything that mathematical logic can describe can be formalized and executed by the machine.
In 1943, a work came out that laid the foundation for all modern wonders that demonstrate neural networks (for example, the ability to distinguish cats from dogs at a superhuman level): "A logical calculus of ideas related to nervous activity" by Warren Pitts. In particular, they proposed a formal model of a neuron. It took a little time, and Alan Turing proposed the famous test of the name of himself - a test to check the creation of artificial intelligence. A year later, in 1951, Marvin Minsky built the first neural network machine (SNARC).
The main problems of artificial intelligence were the lack of computing power of machines and the so-called paradox of Moravec (Hans Moravec): “Why are complex tasks, like proving theorems, simple for a computer, and simple, like recognizing faces, are difficult?”
Until recently, there was no particular interest in AI as a field for research, but Moore’s law, which had been working properly since 1965, made itself felt. The first harbinger of the new spring, perhaps, was the victory of the computer Deep Blue in a chess battle with the then world champion Garry Kasparov in 1997. By the way, in the same year the work of Jürgen Schmidhuber's Long Short-Term Memory (Long Short-term Memory) was published, laying the foundations of the modern generation of recurrent networks.
Further victories in the piggy bank AI became only more. For example, the last well-known victory in the game of go belongs to the program AlphaGo. But even before that, the tread of artificial intelligence was audible: obtaining superhuman quality in the classification of pictures into 1000 classes (ImageNet), speech recognition (ByteNet) and the proof of mathematical theorems (many specialized packages). The new heyday of AI came in the 2000s and continues to this day.
In the 90s, the concept of an “intelligent agent” became increasingly widespread, making decisions about its behavior based on data from the outside world. Attempts to make such agents began in ancient times, but only with the development of sufficient power of computers (we remember about Moore's law) such agents became widespread. Now, intelligent agents control trading on the stock exchanges, land planes and can drive your car for you (Tesla unmanned vehicles immediately come to mind, but ordinary cruise control can also be considered such an agent).
From real-life tasks, we turn to unrealistic tasks, namely, computer games. And then Demis Hassabis (Demis Hassabis), an AI researcher, game developer and chess player, enters the scene. In 2010, he founded the company DeepMind. Demis, being a game developer, retained an interest in them and doing artificial intelligence. So, in 2012, the first results of intellectual agent wins in games for Atari consoles appeared. In the spaces of the former USSR, these consoles have not become widespread, but the Dendi prefix has similar computational capabilities and many Atari ported games, such as Galaxy / Space Invaders. In 2014, the company was acquired by Google, but Demis retained his leadership position and continued to move in the same direction. In 2015, the next milestone was a publication in the journal Nature . Google also found the use of the results obtained in reality - for example, for managing the power supply of its data centers .
We have already mentioned Tesla, it is worth mentioning its founder, Ilona Musk, who, among other things, in 2015 became one of the co-founders of the OpenAI institute, whose goal is to create artificial intelligence. Why does the title have “open”? Ilon Musk postulates that the potential risk of creating artificial intelligence (quite well illustrated in the now classic movie “Terminator” with Arnold Schwarzenegger, filmed in 1984, before the beginning of today's surge of interest in AI) can be reduced if everyone gets equal access to the results of research on artificial intelligence. Not quite an obvious strategy, but we can recall the scientists participating in the Manhattan Project, who transmitted the drawings of the USSR atomic bomb to avoid a dangerous imbalance of armaments.
Already a few months after its creation, in April 2016, OpenAI launched the open-access gym platform (“gym”), where everyone can try out their algorithms to solve intelligent agents in different environments - like moving a cube to the right place in the room or holding the pole in a vertical position. Six months later, Atari games are added to the gym - and now anyone can compete with DeepMind experts in the ability to train intelligent agents to play computer games.
But even before this became possible, the idea emerged to try its hand at the Atari AI competition. Thus was born the first hackathon DeepHack, which took place in the summer of 2015.
Together with the hackathon, it was decided to conduct a scientific school so that the lectures and communication with lecturers would help the hackathon members go further and surpass the results of DeepMind. Looking ahead, we can say that to some extent it was possible: one of the teams even published a scientific paper on the results of the hackathon. Lecturers at the first DeepHack hackathon were: Jürgen Schmidhuber (Juergen Schmidhuber) - it seems that this person needs no introduction, suffice it to say that he is considered one of the fathers of modern neural networks and is the head of the group at the Swiss Institute for Research on AI, Joshua Bengio ( Yoshua Bengio) - another of the fathers of modern neural networks, the head of the AI ​​research laboratory at the University of Toronto, Ruslan Salakhutdinov, the head of the research division of Apple, a professor at Carnegie Mellon University, as well as many others. For a detailed list of participants, lecture notes and slides, please visit game.deephack.me .
Six months later, in the winter of 2016, we decided to switch to a no less interesting task - answers to questions for the American counterpart of the USE. At this time, the Kaggle world competition began with data provided by the Allen Institute for Artificial Intelligence Allen Research Institute (Allen Institute for Artificial Intelligence), founded by Paul Allen (Paul Allen, co-founder of Microsoft) in 2013.
A scientific school also worked here, where leading scientists in the field of natural language processing were invited: Christopher Manning, professor at Stanford University, author of the classic book “Basics of natural language statistical processing”, Tomas Mikolov, researcher at Facebook by Word2vec, Phil Blunsom, a professor at Oxford University, a researcher at DeepMind. And again, many more - details on qa.deephack.me .
For the new DeepHack, which will be held February 6-12, 2017, OpenAI gym was chosen as a public platform, and the game is complex from the point of view of the machine, such as Pacman (we remember the Moravek paradox). Within the framework of the hackathon a scientific school will again work, anyone can attend lectures. Lecturers will be announced later, all updates can be tracked at rl.deephack.me . But now the qualifying round for participation in the hackathon is announced: you need to achieve the best results in the game Skiing on Atari. The first 8 teams will be selected to participate, the results of the qualifying round will be announced on January 22, 2017. We will be glad to see you at the hackathon and scientific school!
And if you are not completely in the subject line and, for example, it is not clear to you why computer games, how they relate to the management of data centers and what will actually be in February, then immediately go under the cat - there is a maximum immersion in the life of an artificial intelligence from antiquity to the present day. Well, you do not think that all this was invented only in the XXI century?
Cheat sheet to the reader: DeepHack is a format that combines a classic hackathon with pizza, Coca-Cola and round-the-clock coding and a scientific school with lectures from leading experts on the topic. What was told on past hackathons and what results turned out can be found on game.deephack.me and qa.deephack.me .
')
Introduction
Delve into history. As you know, the first attempts to create artificial intelligence began a long time ago. The answer to the question, how long ago, strongly depends on what we are putting into the concept of “artificial intelligence”. We will assume that artificial intelligence is some kind of machine capable of making decisions using logical calculations. Such a definition will allow us to miss all (certainly very interesting) attempts to create artificial intelligence in Antiquity, Antiquity, the Middle Ages and the New Time, immediately skipping to the 19th century to Charles Babbage and Ada Lovelace. They are widely known as the creator of the computer and the first programmer, respectively. But let's not linger here and go even further - to the beginning of the 20th century, when David Gilbert, who became Henri Poincaré after the death as the universally recognized mathematician No. 1, put to the world mathematical community the question: “can all mathematical reasoning be formalized?” other well-known scientists gave this question: Kurt Godel with his incompleteness theorem (1931), Alan Turing, who proposed the car named after himself (1936) and Alonzo Church, who invented the λ-calculus (series of works in the 1930s). Gödel's work demonstrated that mathematical logic has its limits. But the results of Turing-Church showed that everything that mathematical logic can describe can be formalized and executed by the machine.
In 1943, a work came out that laid the foundation for all modern wonders that demonstrate neural networks (for example, the ability to distinguish cats from dogs at a superhuman level): "A logical calculus of ideas related to nervous activity" by Warren Pitts. In particular, they proposed a formal model of a neuron. It took a little time, and Alan Turing proposed the famous test of the name of himself - a test to check the creation of artificial intelligence. A year later, in 1951, Marvin Minsky built the first neural network machine (SNARC).
The main problems of artificial intelligence were the lack of computing power of machines and the so-called paradox of Moravec (Hans Moravec): “Why are complex tasks, like proving theorems, simple for a computer, and simple, like recognizing faces, are difficult?”
Revival of interest in AI
Until recently, there was no particular interest in AI as a field for research, but Moore’s law, which had been working properly since 1965, made itself felt. The first harbinger of the new spring, perhaps, was the victory of the computer Deep Blue in a chess battle with the then world champion Garry Kasparov in 1997. By the way, in the same year the work of Jürgen Schmidhuber's Long Short-Term Memory (Long Short-term Memory) was published, laying the foundations of the modern generation of recurrent networks.
Further victories in the piggy bank AI became only more. For example, the last well-known victory in the game of go belongs to the program AlphaGo. But even before that, the tread of artificial intelligence was audible: obtaining superhuman quality in the classification of pictures into 1000 classes (ImageNet), speech recognition (ByteNet) and the proof of mathematical theorems (many specialized packages). The new heyday of AI came in the 2000s and continues to this day.
The key task of AI is behavior management.
In the 90s, the concept of an “intelligent agent” became increasingly widespread, making decisions about its behavior based on data from the outside world. Attempts to make such agents began in ancient times, but only with the development of sufficient power of computers (we remember about Moore's law) such agents became widespread. Now, intelligent agents control trading on the stock exchanges, land planes and can drive your car for you (Tesla unmanned vehicles immediately come to mind, but ordinary cruise control can also be considered such an agent).
History DeepMind and Atari, as a sandbox for AI
From real-life tasks, we turn to unrealistic tasks, namely, computer games. And then Demis Hassabis (Demis Hassabis), an AI researcher, game developer and chess player, enters the scene. In 2010, he founded the company DeepMind. Demis, being a game developer, retained an interest in them and doing artificial intelligence. So, in 2012, the first results of intellectual agent wins in games for Atari consoles appeared. In the spaces of the former USSR, these consoles have not become widespread, but the Dendi prefix has similar computational capabilities and many Atari ported games, such as Galaxy / Space Invaders. In 2014, the company was acquired by Google, but Demis retained his leadership position and continued to move in the same direction. In 2015, the next milestone was a publication in the journal Nature . Google also found the use of the results obtained in reality - for example, for managing the power supply of its data centers .
Transition to open science
We have already mentioned Tesla, it is worth mentioning its founder, Ilona Musk, who, among other things, in 2015 became one of the co-founders of the OpenAI institute, whose goal is to create artificial intelligence. Why does the title have “open”? Ilon Musk postulates that the potential risk of creating artificial intelligence (quite well illustrated in the now classic movie “Terminator” with Arnold Schwarzenegger, filmed in 1984, before the beginning of today's surge of interest in AI) can be reduced if everyone gets equal access to the results of research on artificial intelligence. Not quite an obvious strategy, but we can recall the scientists participating in the Manhattan Project, who transmitted the drawings of the USSR atomic bomb to avoid a dangerous imbalance of armaments.
Already a few months after its creation, in April 2016, OpenAI launched the open-access gym platform (“gym”), where everyone can try out their algorithms to solve intelligent agents in different environments - like moving a cube to the right place in the room or holding the pole in a vertical position. Six months later, Atari games are added to the gym - and now anyone can compete with DeepMind experts in the ability to train intelligent agents to play computer games.
History hackathons DeepHack
But even before this became possible, the idea emerged to try its hand at the Atari AI competition. Thus was born the first hackathon DeepHack, which took place in the summer of 2015.
Together with the hackathon, it was decided to conduct a scientific school so that the lectures and communication with lecturers would help the hackathon members go further and surpass the results of DeepMind. Looking ahead, we can say that to some extent it was possible: one of the teams even published a scientific paper on the results of the hackathon. Lecturers at the first DeepHack hackathon were: Jürgen Schmidhuber (Juergen Schmidhuber) - it seems that this person needs no introduction, suffice it to say that he is considered one of the fathers of modern neural networks and is the head of the group at the Swiss Institute for Research on AI, Joshua Bengio ( Yoshua Bengio) - another of the fathers of modern neural networks, the head of the AI ​​research laboratory at the University of Toronto, Ruslan Salakhutdinov, the head of the research division of Apple, a professor at Carnegie Mellon University, as well as many others. For a detailed list of participants, lecture notes and slides, please visit game.deephack.me .
Second DeepHack
Six months later, in the winter of 2016, we decided to switch to a no less interesting task - answers to questions for the American counterpart of the USE. At this time, the Kaggle world competition began with data provided by the Allen Institute for Artificial Intelligence Allen Research Institute (Allen Institute for Artificial Intelligence), founded by Paul Allen (Paul Allen, co-founder of Microsoft) in 2013.
A scientific school also worked here, where leading scientists in the field of natural language processing were invited: Christopher Manning, professor at Stanford University, author of the classic book “Basics of natural language statistical processing”, Tomas Mikolov, researcher at Facebook by Word2vec, Phil Blunsom, a professor at Oxford University, a researcher at DeepMind. And again, many more - details on qa.deephack.me .
New DeepHack
For the new DeepHack, which will be held February 6-12, 2017, OpenAI gym was chosen as a public platform, and the game is complex from the point of view of the machine, such as Pacman (we remember the Moravek paradox). Within the framework of the hackathon a scientific school will again work, anyone can attend lectures. Lecturers will be announced later, all updates can be tracked at rl.deephack.me . But now the qualifying round for participation in the hackathon is announced: you need to achieve the best results in the game Skiing on Atari. The first 8 teams will be selected to participate, the results of the qualifying round will be announced on January 22, 2017. We will be glad to see you at the hackathon and scientific school!
Source: https://habr.com/ru/post/318758/
All Articles