On October 19th, the British "Nature" magazine published a paper. A new version of the "AlphaGo" computer program can start from a blank state, without the need for any human input. .
The creator named it "Zero". The Taoist said that Daosheng had one, two, three and three births. Starting from scratch, this can surpass the Go skills of the top human players. It sounds subversive and makes many people fear.
In addition, Ke Jie, a world-famous go player defeated by AlphaGo, said on the social platform soon: “A pure, purely self-learning AlphaGo is the strongest... For the self-advancement of AlphaGo... There are too many human beings (please Contact the context, note the context."
The concept of "human excess" is taken out of context. An algorithm can learn superb chess skills only by itself. Artificial intelligence can "zero" everything and thus be interpreted. However, is this the case? Science and Technology Daily reporters interviewed industry experts and listened to what they said.
Listening: Zero also needs a database
"Zero can generate its own data. The next second's data and the last second's data are played against each other. If it wins, it will be generated again. This way, the winning move will be gradually 'evolved' out." China Chief Data Officer Coalition Gao Yuzhao, a member of the expert group and founder of Kesi Technology, said.
Then, without database support, how does Zero generate data and how does it know what kind of data it produces? Is there really "smartness?"
Hua Dai Gene CEO Yin Hao does not agree with the "self-taught" argument. He compared AlphaGoZero and AlphaGo on the source of learning. "The latter was entered into the 30 million chess games in human history, learning the algorithm for Go rules step by step, and Zero is standing on AlphaGo's shoulders, inheriting the latter's well-defined algorithm to understand Go. The rules, following this rule, beat each other."
In other words, Zero learns rules from his predecessors. "It doesn't need only the data of past human races, and it doesn't mean that it doesn't require a database," Gao said.
Professor Yan Endong of Beijing Language and Culture University interprets Zero as "a chess player" with two reasons. First, "Under the same rules, the machine's catching strategy is different from what one can catch." The second is "It The speed of operations is faster than humans, so it is possible to perform more in-depth or even exhaustive calculations and know exactly what to do."
In other words, it is not only looking right, but also running fast. However, there are preconditions for such a situation - "The rules are simple! The outcome is clear! Strategies can be exhaustive!" Yan Endong said.
Analyze Zero's learning process to understand why it must be such a problem before they can win.
According to the rules, it constantly generates new data, and then the old and new data compete against each other and eventually produce a winning or losing result. In other words, under the fixed rules, the "winning and losing" is constantly verified, so that Zero gets a fine chess game, and when all these strategies are verified once, it is invincible.
In 1997, the computer "dark blue" defeated the famous chess player Kasparov. "This incident also caused a sensation at the time," said Yoon Yeung. "At that time, chess could not be exhausted. However, with the advancement of hardware operation speed and the improvement of the algorithm, chess can be exhausted by 2005. This means that no matter which step you take, the computer can take a step-by-step look at N-steps, calculate all possibilities, and then give all the corresponding methods. With the application of quantum computing, the exhaustive move of Go moves is also possible.
"In terms of algorithms, Zero's use of confrontational reinforcement learning is the latest advancement in machine learning algorithms. It has important theoretical and practical implications for the development of artificial intelligence," said Yan Endong.
Positioning: Still in the computational intelligence phase
"Actually, Go is a simple 'calculus intelligence'. The reason why it is so focused is because Go has always been a symbol of strategy." Yan Endong said.
“Achieving a winning streak and winning a thousand miles away†can always be reminiscent of such a scene. In the account of the Chinese army, a smoky haze is lingering, chessboard pieces are set, the account is a black and white game, and the accounts are thrown into battle. .
Go is a symbol of wisdom. When human beings are unable to conduct full-sample analysis, how to select the optimal strategy based on some samples reflects the wisdom of judgment. When the computing power can calculate the full sample, the type of intelligence changes -
"Through excellent computing power, Zero can have a fast convergence (accurate trend) according to the algorithm, and can achieve near-globally optimal results according to the strategy." Yan Endong said, which means that it does not need to judge the choice but needs to find it tirelessly. Best solution, keep trying. "Because the calculation depth is deeper, an optimized parameter calculation strategy has been adopted, the optimization process has been accelerated, and a better calculation model has been obtained."
“So far, the AI ​​of landing applications has evolved from speed, automation, ease of deployment, and so on—speed evolution is based on hardware upgrades, distributed processing, etc.; automation means that without manual tagging, AI can automatically select useful information. Conducting memory training is also part of easy deployment.†Takahashi said that such “evolution†is the reason why Zero uses four TPUs (professional chips needed for neural network training) and tries to surpass his predecessor with 4.9 million chess games.
However, it is no better at dealing with complex issues than humans. Remember the endless exams when I was young? The multiple-choice questions and the judgment questions AI with more definite answers can be done, and the essay questions are almost the same.
Yan Endong has given a more professional classification. “The development of artificial intelligence is divided into three levels: computational intelligence that can survive and count, perception intelligence that has audio-visual touch, and cognitive intelligence that can understand and think. Zero is still at the stage of computational intelligence. â€
Yin Hao also believes that the database-independent bi-pulse algorithm cannot cope with unclear calculations. "For example, in the medical and health industry, data is still the king. The algorithm will be revised constantly according to the accumulation of data, and it will move from artificial intelligence (AI) to real. Intelligent (RI)."
Outlook: More advanced intelligence is still on the way
Yan Endong gave a witty example. "For example, when the word "your sister" is used to refer to a person and when an emotion is expressed, it is very difficult to make AI clear."
Yin Hao also believes that there are two kinds of languages. Languages ​​that exist only for communication do not need to be learned, but language learning to express thoughts and emotions will receive more attention.
In fact, the research that allows AI to acquire perception, cognitive intelligence, etc. has been ongoing. “Cognitive intelligence currently needs to go a long way,†said Yan Endong.
Data shows that 55% of technology companies in the AI ​​industry are in computer vision and 13% in natural language processing. Tactile research has also been incorporated into the national key R&D program.
Perceived smart vision, hearing, touch, being in industry research, basic research and other fields continue to expand. Yan Endong said, “Language intelligence is an important research direction of artificial intelligence, and Chinese syntax and semantic analysis is the core technology of language intelligence. Computers lack sufficient features to capture Chinese contextual language information, and no substantial breakthroughs have been made in the accuracy of analysis.â€
In order to allow Chinese to be included in the AI ​​language system, the Center for High-level Linguistic Resources at Beijing Language University conducts semantic analysis in order to obtain Chinese syntax syntax with high robustness, high accuracy, and linear velocity computational complexity. Analyzer. "Robustness is to let the machine have the ability to deal with various forms of language, including non-standard expression." Yan Endong said, "We have established Chinese analysis of big data above TB level, the future AI can understand written language, spoken language, pun, Joke……"
Visually, the Shanghai Fire Research Institute of the Ministry of Public Security and other units have already been able to see the flames sending fire alarms. In terms of touch, inaccurate interactive information such as gestures, postures, touch, voice, expression, eye movement, and physiology is being captured, recognized, understood, and even integrated.
"These are not just end-to-end issues, but problems that require complex strategies to solve. The AI ​​system requires multiple inputs, and also expects multi-factor output." Yan Endong said that in these areas, the human experience database cannot be Put aside, "Just like building a wall, Zero can be seen as building a new wall, and the higher stage is to make up for an incomplete wall."
Amd Gaming Pc,Gaming Computer,Gaming Pc I7,Pc Gamer
Shenzhen Innovative Cloud Computer Co., Ltd , https://www.xcypc.com