Imitating human logic, artificial intelligence passed the second year of science exams, the researchers: completed the boss’s wishes


Editor’s note: This article comes from WeChat public account “Big Data Digest” (ID: BigDataDigest), compiled beer bubble, oak _Hiangsug, 36 氪 authorized release.An artificial intelligence called Aristotle has just passed the scientific test of the eighth grade in the United States. Last week, this news occupied the first edition of many US news websites.The eighth grade of the United States is probably equivalent to the second day of the country. How difficult is the science test for the children in the second day?To answer this question, let’s first take a look at two American eight-year scientific test multiple-choice questions.1. The organization in the human body that can work together to accomplish a specific function is called: organismC.a systemD.a cell2. Which of the following changes is most likely to result in a decrease in the number of squirrels in a certain area?A. The number of predators is reduced. B. The competition within the squirrels is reduced. C. The amount of food available is reduced. D. The increase in the number of forest fires. Obviously, these two questions belong to two different types.The first question belongs to the knowledge point, and it can be answered as long as it is carefully carried forward; the second one is a logical reasoning problem.Most children may be more willing to answer the second inquest of this logical inference, but for artificial intelligence, the situation may be just the opposite.AI is the eighth grade paper, and the correct rate of multiple-choice questions is over 90%. On Wednesday, the famous laboratory in Seattle, “Allen Artificial Intelligence Institute”, released a new article called “Aristo”.The artificial intelligence system, which correctly answers more than 90% of the eighth grade scientific test questions, and achieved more than 80% accuracy in the twelfth grade exam.This artificial intelligence through testing capabilities shows that researchers have made tremendous progress in a few months, and artificial intelligence systems can understand the language and simulate human decision logic.Aristo’s setting is only used to answer multiple choice questions.It took several standard exams for New York candidates, but the Allen Institute removed the questions that included pictures and graphics. Answering these questions requires additional skills—the ability to combine language understanding with computer vision logic.Some test questions only require some information extraction capabilities, such as the first question above, which is good at artificial intelligence.However, scientific testing is not the kind of thing that can be done by simply remembering the rules. It needs to use logic to establish connections.For example, the second question, the increase in the number of forest fires directly leads to the death of squirrels, or the reduction in the source of food makes them unable to multiply.Artificial intelligence needs to understand such logic in order to answer the correct question.In fact, before Aristo’s success, AI has been hanged countless times.In 2016, more than 700 computer scientists participated in a challenge of 80,000 US dollars (about 570,000 yuan), titled “eight-year science test” – but the answerers were not these scientists, but they establishedArtificial intelligence system.Unexpectedly, the candidates were completely “hanging”, and even the most mature artificial intelligence system could not answer more than 60% of the questions. The language level and logic level were far behind the eighth grade students.Behind Aristo’s Aristo is Bert 2016. When AlphaGo defeated human professional Go player Li Shishi, many people think that the turning point of artificial intelligence is coming.However, the excitement of Dr. Oren Etzioni, a former professor at the University of Washington and the current technical director of the Allen Institute of Artificial Intelligence, quickly subsided.He said that artificial intelligence is not as advanced as it seems.He mentioned the game that the Allen Institute had participated in before, and an eighth-grade scientific test made it difficult to build an artificial intelligence system.The Allen Institute quickly improved its previous work and set about building Aristo faster than many experts, including Dr. Etzioni.Aristo’s ability to test comes from neural networks. In recent years, the world’s top artificial intelligence laboratories, such as Google, Facebook and other companies’ laboratories, have used neural networks for natural language processing (NLP), which can analyze human articles.And books to learn the complex changes in the language.At the end of last year, the Google AI team released the BERT model, which showed amazing results in the machine-reading comprehension level test SQuAD1.1: all two metrics surpassed humans and created the most in 11 different NLP tests.Good results include pushing the GLUE benchmark to 80.4% and the MultiNLI accuracy to 86.7%.The full name of BERT is Bidirectional Encoder Representation from Transformers, which is the Encoder of two-way Transformer. The main innovation of the model lies in the pre-training of the model. The methods of Masked LM and Next Sentence Prediction are used to capture the statement respectively.The Bert model architecture Dr. Etzioni quickly realized that the Aristo system could be built on top of Bert, and they used the Bert model to train a wide range of question and answer data.Aristo uses eight types of agents to answer questions based on different types of topics—including agents looking up answers in the database, agents checking related concept lists, and agents performing qualitative reasoning.Each agent will have a probability of correctness for multiple choice answers, and Aristo will weight the probability of different options to select the most likely one or more. The model is optimized through multiple rounds of training and calibration.For example, one question is: How is the iron atom in the iron block affected when the block melts?A. Iron atoms increase the quality.B. Iron atoms contain less energy.C. Iron atoms move more frequently.D. The volume of iron atoms increases.To answer this question, Aristo first looked for the knowledge that “iron atoms move faster as heat increases”, linking the term “melting” to “heat”, linking the term “fast” to “frequent” and CRating is the correct choice.Combining different problem-solving methods has cleared Aristo’s test scores from about 60% in 2016 to 91.6% this year.In the 12th grade exam, the model score was 83.5%.Is Aristo’s constantly improving answer accuracy rate learning or scum?It can be used!Some scientists don’t have much enthusiasm for the progress Aristo has made. They think that the machine still has a long way to go to master the natural language, not to mention thinking like a human student.”We can’t compare this technology to real students and their logical reasoning skills.” Jingjing Liu, a researcher at Microsoft who is involved in several similar technology developments, said.Liu and her Microsoft colleagues have tried to establish a system that can pass the GRE exam – GRE is a mandatory test for graduate admissions in the United States.Liu said that dealing with the language part is feasible, but building logical reasoning skills that can be used to deal with mathematical problems is another matter.“This is really a too challenging job.” But from a business perspective, this evolution of Aristo will have a wide impact on many products and services, from Internet search engines to hospital documentation systems.According to the New York Times, Dr. Etzioni said: “This technology will bring important business results. I can confidently say that you will see a new generation of products brought about by this progress, possibly from startups.It may come from a big company.” “This technology is still in its infancy,” said Jeremy Howard, technical director at “But the potential of its technology is limitless, and we are far from fully exploiting the potential of this technology.”OMT, Aristo is also the founder of the Allen Institute. The Allen Institute is named after Paul Allen, co-founder of Microsoft. He founded the Allen Institute of Artificial Intelligence in 2013, hoping to solve the problem.A major issue of intelligent development.The artificial intelligence science challenge, with the “eight-year science test” as the title, stems from a selfishness of the Seattle billionaire: he wants researchers to design an artificial intelligence program that is smart enough to pass the eighth-grade science test..Since its inception, researchers at Allen Research have been working on building this smart artificial intelligence program, Aristo.This is not an easy task. The researchers have tried countless times in the past five years, but they have not achieved the effect that Allen hopes.However, in October last year, it was not yet time to witness the birth of Aristo. At the age of 65, Allen passed away.In different emails, Aristo’s authors Etzioni and Clark paid tribute to Paul Allen.When asked if such a system Allen could be satisfied, both said: “No.” Etzioni and Clark at the Allen Institute of Artificial Intelligence “Paul will be very happy, but will not let us be satisfied with the presentSome honors,” Etzioni said. “He will ask: What is the next important stage of NLP?” “I can imagine he would say ‘Congratulations! But what is the next step?” Related reports: