Anecdotes of Watson supercomputer before the game

From June, 5438 to February, 2009, Friedman and other producers saw Watson for the first time, when the supercomputer confronted two human players. "Watson" uses IBM POWER7 server, which can optimize the tasks it handles. IBM said that Watson, who participated in the challenge, had to deal with all tasks quickly. In addition, this supercomputer also adopts a series of patented technologies, which can perform task and data processing and real-time information analysis at the same time.

Farusi pointed out that the essence of danger is to make technology develop in the right direction. He said: "This program covers a wide range of fields and will raise all kinds of questions, which is one of the challenges we hope to participate in. This challenge is a test of players' confidence. You must be sure that the answer is correct before you can answer. In addition, you have to give an answer quickly. "

IBM said that the technology used by Watson can play a role in many fields, such as medical care, which can help improve the accuracy of doctors' diagnosis of diseases, improve online self-service desks, provide tourists and residents with specific information about relevant cities, or provide support to consumers through mobile phones. In order to prepare for danger, "Watson" has fought the former dangerous champion for more than 50 times. In addition, I took part in the test prepared by the program group Danger for all potential players and passed it successfully.

Harry Friedman, the producer of Danger, said that when IBM first came into contact with Danger, the producers were very interested, but they were also worried that the audience would disapprove of Watson's participation in the competition and regard it as a gimmick or trick. He said: "The attitude is different now. This is a contest between the knowledge gained by computers and the knowledge gained by the best dangerous players. This is an important game and we all want to participate. " Watson consists of 90 IBM servers and 360 computer chip drivers. It's a computer system the size of a 10 ordinary refrigerator. It has 15TB memory, 2880 processors and can perform 80 trillion operations per second (this is the current situation). These servers use the Linux operating system. IBM's processor for Watson is Power 7 series processor, which is the strongest processor in RISC (Reduced Instruction Set Computer) architecture. Manufactured by 45nm technology, eight nuclear has 32 threads, the highest frequency can reach 4. 1GHz, and the second-level cache can reach 32MB. It stores a large number of books, news and movie scripts, dictionaries, anthologies and the World Book Encyclopedia and other millions of materials. Every time he reads the prompt of the question, Watson digs three feet in his database in less than three seconds, searching through 200 million pages of data.

Watson is developed based on IBMDeepQA (Deep Open Domain Question Answering System Engineering) technology. As the foundation of Watson supercomputer, DeepQA technology can read millions of pages of text data, use deep natural language processing technology to generate candidate answers, and evaluate those questions according to many different scales. More than 100 groups of algorithms developed by IBM R&D team for Watson can analyze problems, retrieve millions of pieces of information in 3 seconds, then filter and restore them to answers and output them to human language. Each algorithm has its own special function. One of the algorithms is called nested decomposition algorithm, which can decompose clues into two different search functions.

Who is smarter, Watson supercomputer or Google search engine? For this problem, Steven Becker, the former science and technology editor of Businessweek, recently concluded through analysis and comparison that Watson supercomputer is smarter than Google. Becker thinks the answer is simple, and Google can't answer this question yet. Google's dependence on our human brains is manifested in two aspects: First, when describing users' query requests, Google will make users think like computers, select three or four messages that best meet the computer's intentions, and form a query result list. Secondly, Google will guide users to find answers similar to their own queries, so that users can spend more brain power to find the exact answers they need to query. Watson supercomputer handles all the problems by itself. It must decode complex English, exhaust all possible answers, choose one from them, and finally decide whether it meets the requirements. For computer Watson, it is an arduous task to choose the appropriate context, especially in games full of hints and pranks. It must also determine the reliability of the answers and whether it is necessary to take risks and guess. Watson needs to recognize human language and analyze subtle meanings, such as satire, riddles, word formation, poetic clues, etc. Through a series of data comparison and simulation of human association ability, he can get accurate answers, and then answer them in human language. In addition, Watson can imitate human beings to think strategically about problems, such as the choice of problems, whether to skip topics that he is not good at, and even imitate human tone to be funny. The whole process is completed only by computers and software algorithms, without the participation of engineers (this kind of quick question and answer is actually impossible), and the network is also disconnected, so the computer can't cheat and must think for itself.

Dr David Ferrucci, a scientist who led the IBM research team that designed Watson, said: After four years of hard work, our scientific team believes that Watson has been able to quickly understand the content of the "dangerous edge" question, analyze the information it needs, get accurate answers and give credible answers. Watson will estimate the remaining bonus amount of the program group "Danger Edge", how much he is behind or ahead of other competitors, his performance on specific topics, and adjust his information level in time. If Watson lags behind his opponent too much, his confidence level will be low, such as 40%, and he will continue to work hard at this time; However, when it is ahead of its opponent, it will also think: I have been ahead so much, why take the risk? Therefore, even if its confidence value is 75%, it may choose not to answer.

During the four years of developing Watson, the R&D team mainly evaluated the system in two ways. First of all, they conduct large-scale tests in batches (for example, 3000 questions at a time), evaluate system performance, implement error analysis, and improve system performance. Based on this multi-problem performance, important performance evaluation can be made from the statistical point of view; Secondly, the second way for the team to evaluate Watson is to practice with the former contestants of Danger Edge. In the winter of 2009, they played 79 games with players who had appeared in Danger Edge. In the past autumn, Watson played 55 games with the player who won the championship in Danger Edge. These sparring competitions gave the R&D team a deep understanding of Watson's performance. Danger Edge is a quiz show of CBS, which has gone through decades of history. The competition of this program is conducted in a unique question-and-answer form, and the questions set cover a wide range, involving history, literature, art, popular culture, science and technology, sports, geography, word games and so on. According to the clues provided in the form of answers, participants must give short and correct answers in the form of questions. Contrary to the general question-and-answer program, the danger edge asks questions in the form of answers and answers in the form of questions. Participants need to have knowledge of history, literature, politics, science and popular culture, as well as the ability to analyze obscure meanings, satires and riddles, and computers are not good at such complex thinking.

Watson's original intention was to understand more complex words, languages and human knowledge. David Ferrucci, an IBM scientist and head of Watson's team, said that using dangerous edges to develop computer systems will push technology in the right direction. He said that this program will ask all kinds of questions. This also involves self-confidence. Don't answer when you think your answer is incorrect. You still need to make a quick judgment. IBM said that winning the "dangerous edge" competition is not the main purpose. Through Watson's technology, doctors can diagnose cases faster and lawyers can study cases faster. These are very important and we want to be a part of them, said Harry Friedman, executive producer of Danger Edge. On 2011February 17 Beijing time, the final result of man-machine war was released: computer Watson defeated human beings. Watson, a supercomputer jointly developed by IBM and the University of Texas, beat ken jennings and brad rutter, the two most successful contestants in the history of the most popular quiz show in the United States, and became the king of the new show.

On the third day of the competition, Watson, IBM's supercomputer, scored 465,438+0,465,438+03 dollars, while ken jennings and brad rutter, two human athletes, only got 654,338+0.92 million dollars and 654,380+0.02 million dollars respectively.

Add up the results of the three match days and get the final total score. Watson is also far ahead of humans. In the final score, Watson reached $77 147, ken jennings ranked second, but only got $24,000, and Brad Ratt got $2 1600, ranking third.

On the third match day, Watson led all the way, so that before entering the final Jeopardy, there was almost no chance for human players to surpass Watson.

In the final "Jeopardy", ken jennings, the second player, had given up chasing Watson and chose to keep the second position, so he only gambled $65,438+$0,000, while Brad Ratt, the third player, risked the biggest bet he could make-5,660. Watson once again exposed his inhumanity and bet $65,438 +0.7973. Columbia University Medical Center and University of Maryland School of Medicine have signed contracts with IBM, and the medical staff of the two universities will use Watson to diagnose and treat diseases more quickly and accurately. There are many professional papers published in journals in its huge information base, which enables doctors to use the latest scientific research results to treat patients.

Watson needs to improve to really become a doctor's right-hand man. Doctors need more than just answers. Moreover, sometimes the information provided by patients is inaccurate or contradictory, which requires doctors to make judgments with rich experience. The next challenge for the IBM R&D team is to get Watson to provide some hypothetical situations. It will take the R&D team at least two years to complete this task. ? 14 years ago, the computer "Deep Blue" developed by IBM defeated Kasparov, the chess champion. Now, the computer named after its founder Thomas J. Watson continues to challenge the limits of human intelligence.

IBM is good at making all kinds of complicated machines. In addition to servers, there are intelligent computers (minicomputer clusters). It seems that IBM is more than just a company that produces philosophy, just like the ironic buzzword says.

In the1960s, after the technical research and development of artificial intelligence stagnated for several years, scientists found that if artificial intelligence was defined by simulating the human brain, it would enter a dead end. Now, the latest definition of artificial intelligence is "to complete the task of decentralization through machine learning, large-scale databases, complex sensors and ingenious algorithms", which has replaced the once popular "rebuilding the brain".

According to this definition, Watson is considered to have taken another step in artificial intelligence. "Deep Blue is only doing super-large-scale calculations, and it is the embodiment of human mathematical ability," Pan Yue, senior manager of IBM China Research Institute, told CBN Weekly. He also participated in the Watson Project and was responsible for providing data support. "In the fields of machine learning, large-scale parallel computing, semantic processing, etc., Watson's greatness lies in integrating these technologies into an architecture to understand human natural language."

If these explanations are a little obscure, you can refer to the rules of the game in Danger Edge, and you will know the value of Watson.

On the brink of danger is a long-lasting TV question-and-answer program of CBS, which started broadcasting at 1964. The most wonderful thing is that the problems in the game are all-encompassing, covering almost all fields of human civilization. Its rule is that you can get a bonus if you answer the question correctly, and deduct the money if you answer it incorrectly.

For humans, the rules are simple, but for Watson, it means many challenges. First of all, Watson must understand the natural language of the host, which deep blue does not have; Secondly, Watson needs to analyze these languages, such as irony, puns and conjunctions, and then judge the meaning of the topic according to the keywords, and Watson conducts related searches to evaluate the possibility of various answers; Finally, choose the three most likely answers. When one of them exceeds 50%, the program starts and Watson presses the transponder.

All these are realized by 90 IBM servers, 360 computer chip drivers and DeepQA system developed by IBM. The processors provided by IBM for Watson are Power 750 series processors, which are the strongest processors in RISC (Reduced Instruction Set Computer) architecture-these processors can support Watson to finally get a reliable answer in less than 3 seconds.

Watson focuses on the parallel operation of minicomputers. It is difficult to shake IBM's position in the mainframe field-when Watson Jr, the son of the founder Watson Sr., took charge of IBM, IBM invested about $6 billion, exceeding the R&D expenditure of NASA that year, and finally squeezed out Japan's NEC and France's Cray, establishing IBM's monopoly position in the mainframe market.

"There is a joke circulating in IBM's internal staff training, that is, in the mainframe field, 97% of the market share comes from IBM, and the remaining 3% comes from the eliminated IBM second-hand mainframe." The person in charge in front of the IBM server sales department told CBN Weekly.

IBM tries to bring the advantages of mainframe into the field of minicomputer. Watson focuses on the parallel computing of 90 minicomputers, which is the new sales star of IBM. "The computing speed of IBM minicomputers is outstanding," said the above-mentioned IBM employees, "because it directly transplanted the CPU R&D advantages of mainframes to minicomputers."

IBM is one of the few manufacturers that can independently develop minicomputer CPU, and IBM has also sold these technologies to chip manufacturers other than Intel.

Now in the minicomputer market, the market ratio of IBM and HP is roughly 2: 1, while SUN, another minicomputer company, is now mainly positioned in the low-end market.

In addition, a model of IBM's global R&D team also increased Watson's chances of winning the competition.

These teams have a very detailed division of labor. For example, Haifa team in Israel is responsible for the search process of the deep open domain question answering system project, Tokyo in Japan is responsible for Watson's word meaning and word connection in the question answering, IBM China Research Institute and Shanghai Branch are responsible for providing Watson with data support from different resources, and there are also teams specializing in algorithms and game teams for strategic betting.

"We are like every different virtual force, and everyone only does what they are best at and familiar with." Pan Yue said.

In fact, this is the characteristic of IBM, which has always followed the successful model of the industrial age, such as the assembly line model that manufacturing enterprises are better at, and made its team more efficient through the reward and punishment system of grading employees.

These DeepQA systems developed by R&D genius ensure that Watson can have a brand-new human-computer interaction mode, such as understanding and analyzing natural languages. In fact, understanding natural language has always been the strength of IBM research institutions in artificial intelligence.

This is very similar to the bridge in science fiction movies-recruiting talented people around the world and carrying out secret robot development plans. The difference is that IBM's meaning is obviously more practical.

Previously, the AIX operating system based on deep blue has enabled IBM to obtain a large number of orders in commercial applications and government departments. IBM also hopes to apply Watson's DeepQA system to medical services, consulting and other fields.

"Watson's advantage is to give accurate and reliable answers, so it can provide doctors with more suitable solutions for patients." Pan Yue said, "The application in the medical field will be the most important area of Watson's business."

If the Watson project wants to be implemented in the medical industry, it still needs to face legal problems. An IBM researcher said, "If Watson's diagnosis is wrong and the doctor misheard the diagnosis, then Watson will be in danger of being taken to court by the patient. This is an application problem that IBM is considering."

For IBM, Watson will not only continue to challenge the limits of human intelligence in the future, but also help the company compete with Amazon, Google and Microsoft for the dominance of the commanding heights of future technology.