Last year, many game AI reached a level beyond human beings. Artificial intelligence has not only played complex games such as Texas Hold 'em, StarCraft and Dota2, but also been recognized by top journals such as Nature and Science.
The heart of the machine sorted out the seven hottest studies in the field of artificial intelligence and quantum computing last year. Let's look at it in chronological order:
The first major study appeared in February. Following the release of BERT, a 300-million-parameter language model that refreshed the task record of1NLP, Google OpenAI launched a more powerful model again in February 20 19, and this time the model parameters reached1500 million. This is a large unsupervised language model, which can produce coherent text paragraphs and realize SOTA performance on many language modeling benchmarks. In addition, the model can achieve preliminary reading comprehension, machine translation, question and answer and automatic summarization, without the training of specific tasks.
This model, named GPT-2, is a large-scale language model based on Transformer, which contains 65.438+0.5 billion parameters and is trained on 8 million web data sets. Training GPT-2 has a simple goal: given all the previous words in the text, predict the next word. GPT-2 is a direct extension of GPT model. It is trained on the data amount of more than 10 times, and the parameter amount is also more than 10 times.
GPT-2 shows a series of universal and powerful capabilities, including generating conditional synthesis text with the best quality at present, in which we can feed the input into the model and generate very long coherent text. In addition, GPT-2 is superior to other language models trained in specific fields (such as Wikipedia, news or books) and does not need to use training data in these specific fields. In the tasks of knowledge answering, reading comprehension, automatic summarization and translation, GPT-2 can learn from the original text without task-specific training data. Although these downstream tasks are far from the current optimal level, GPT-2 shows that all kinds of downstream tasks can benefit from unsupervised technology if they have enough (unlabeled) data and computing power.
Finally, based on the large-scale universal language model, it may have a huge social impact, and considering that the model may be used for malicious purposes, when releasing GPT-2, OpenAI adopted the following strategies: only the smaller version and sample code of GPT-2 are released, and the data set, training code and GPT-2 model weight are not released.
The best papers at the machine learning summit will always arouse people's extensive discussion. At the ICML 20 19 (International Machine Learning Conference) held in California in June this year, Common Hypothesis in Unsupported Learning, which was jointly written by ETH Zurich, Max Planck and Google Brain, won one of the best papers. In this paper, the researchers put forward a view contrary to the previous academic prediction: without supervision, it is impossible to have an independent representation (decoupling representation) of any data.
Paper link: une does not match).
Based on these advantages and disadvantages, this study proposes a generalized autoregressive pre-training model XLNet. XLNet can: 1) learn bidirectional context information by maximizing the logarithmic likelihood of all possible factorization orders; 2) Overcome BERT's shortcomings with autoregressive characteristics. In addition, XLNet also integrates the idea of the current optimal autoregressive model Transformer-XL.
Extended reading:
2065438+In July, 2009, Depo AI Pluribus successfully defeated five expert human players in the Infinite Hold 'em Six-Person Match. Pluribus was jointly developed by Facebook and Carnegie Mellon University (CMU) * *, which achieved the task that the predecessor Libratus (cold master) failed to complete. This study has been published in the latest issue of Science.
According to reports, the competitions designed by Facebook and Carnegie Mellon University are divided into two modes: 1 AI+5 human players and 5 AI+ 1 human players. Pluribus has won in both modes. If a chip is worth $65,438+0, Pluribus can win an average of $5 per game and $65,438+0,000 against five human players for an hour. Professional poker players consider these results to be decisive victory advantages. This is the first time that AI has defeated top professional players in a large-scale benchmark competition with more than 2 people (or teams).
In this paper, Pluribus integrates a new online search algorithm, which can effectively evaluate its decision by searching the previous steps instead of just searching until the end of the game. In addition, Pluribus also uses a new self-playing imperfect information game algorithm with higher speed. In a word, these improvements make it possible to train Pluribus with little processing power and memory. The total value of cloud computing resources used for training is less than $ 150. This efficiency is in sharp contrast to other recent artificial intelligence milestone projects, and the training of these projects often requires millions of dollars of computing resources.
The result of Pluribus' self-game is called blueprint strategy. In the actual game, Pluribus uses search algorithm to improve this blueprint strategy. But Pluribus will not adjust its strategy according to the trends observed from its opponents.
In the field of quantum computing outside artificial intelligence, there were also important research breakthroughs last year. From 2065438 to September 2009, Google submitted a paper entitled "Quantum Advantages of Using Programmable Superconducting Processors", which was uploaded from the website of NASA. In the experiment, the researchers first proved the superiority of quantum computer over traditional architecture computer: in the experiment of computing 10000 in the world's first supercomputer summit, Google's quantum computer only took 3 minutes and 20 seconds. Therefore, Google claims to achieve "quantum advantage." Later, the paper appeared on the cover of Nature150th anniversary edition.
This achievement stems from the unremitting efforts of scientists. Google's research in the direction of quantum computing has passed 13 years. In 2006, Hartmut Neven, a Google scientist, began to explore ways to accelerate machine learning through quantum computing. This work promoted the establishment of Google AI quantum team. In 20 14, John Martinis of the University of California, Santa Barbara (UCSB) and his team joined Google and started to build quantum computers. Two years later, the papers of Sergio Boiso and others were published, and Google began to focus on the task of realizing the superiority of quantum computing.
Today, the team has established the world's first quantum system beyond the capabilities of traditional supercomputers, which can perform calculations for specific tasks.
The quantum superiority experiment runs on a 54-qubit fully programmable processor named Sycamore. The processor contains a two-dimensional grid, and each qubit in the grid is connected with the other four qubits. The success of the quantum advantage experiment is attributed to Google's improvement of the dual quantum bit gate with enhanced parallelism, which can reliably achieve recording performance even if multiple gates are operated at the same time. Google uses a new control knob to achieve this performance, which can turn off the interaction between adjacent qubits. This greatly reduces the errors in this multi-connected qubit system. In addition, Google has further improved its performance by optimizing chip design to reduce crosstalk and developing new control calibration to avoid qubit defects.
Although AI didn't beat Serral, the strongest human player, its research paper was still in Nature. At the end of 20 19, 10, DeepMind's paper on AlphaStar was published in this issue of Nature, which is the latest research progress of artificial intelligence algorithm AlphaStar. It shows that AI has reached the top level of StarCraft II against ladders without any game restrictions, and its ranking in Battle.net has exceeded 99.8% of active players.
Looking back on the development of AlphaStar, DeepMind announced in 20 17 that it began to study artificial intelligence-AlphaStar, which can play the real-time strategy game StarCraft II. 20 18 12 10, AlphaStar beat Dani Yogatama, the strongest player in DeepMind. 65438+February 65438+February 02, AlphaStar has been able to beat professional player TLO 5:0 (TLO is a zerg player, according to game commentators, its performance in the game can be around 5000 points); Another week later,16,5438+February 9, 2009, AlphaStar also defeated professional MaNa with a score of 5:0. At this point, AlphaStar has taken another step forward and reached the top level of mainstream e-sports games.
According to the description in Nature, DeepMind uses common machine learning techniques (including neural network, self-game with reinforcement learning, multi-agent learning and imitation learning) to learn directly from game data. The gameplay of AlphaStar is impressive-this system is very good at evaluating its strategic position and knows exactly when to approach and when to stay away from its opponent. In addition, the central idea of the paper is to extend the fictional self-game in the game environment to a group of agents, that is, "alliance".
The core idea of the concept of alliance is that winning is not enough. On the contrary, the experiment needs the main agent to win all the players, and the main purpose of the "exploiter" agent is to help the core agent expose the problems and become more powerful. This does not require these agents to improve their winning percentage. By using this training method, the whole agent alliance learned all the complicated strategies in StarCraft II in an end-to-end and fully automated system.
20 19, there are many technological breakthroughs in all directions in the AI field. In the new year, we expect more progress.
In addition, Machine Heart launched its new product SOTA model at the end of September 20 19. Readers can find SOTA papers in corresponding fields and tasks of machine learning according to their own needs, and the platform will provide papers, models, data sets, benchmarks and other related information.