GABRIEL BITAR E NANA LAHOZThe idea of the existence of intelligent computers that can reason, learn and make decisions autonomously has long aroused people’s curiosity and been portrayed in fiction. The most famous of these machines is, without doubt, the HAL 9000, a character endowed with advanced artificial intelligence, immortalized in the movie “2001 A Space Odyssey”, directed by Stanley Kubrick in 1968. However, will these machines be one day able to leave the fictional universe and become a reality? Various researchers around the world, are trying to answer this question with projects to make computers smarter. One is “Read the Web”, a program developed jointly by researchers at the Federal University of São Carlos (UFSCar) and Carnegie Mellon University, in the US The group is developing a computer capable of learning autonomously and using knowledge already acquired to evolve its own learning. The problem proposed is to create a machine that can read web pages and in this way continuously improve the learning capacity of the computer.
“We want to show that with the current development of machine learning techniques, information retrieval and language processing it is possible to build an ‘intelligent’ computer with the capacity to acquire more and more knowledge, exactly like we do,” explains Professor Estevam Rafael Hruschka Junior, from the Department of Computing at UFSCar, and the Brazilian coordinator of the initiative. To achieve this objective, the team built a computer program, called NELL (acronym for never-ending language learner), which is looking for a new standard of machine learning called ‘endless learning.’ “It’s a new paradigm. The computer learns continuously and not only a specific kind of knowledge, but also general knowledge and common sense that will help expand its ability to learn as time goes by,” Hruschka explains.
An important aspect of endless learning is the accumulation of experience. Just as we humans learn concepts that are more complex after we acquire more basic and simple knowledge – it is easier to learn algebra after learning basic arithmetical operations – NELL draws upon its accumulated experience in the process of future learning. “In the same way that an older employee in a company identifies situations that can lead to mistakes and avoids them, over time NELL identifies not very successful learning strategies and can change them in such a way as to improve the learning process, “says the researcher.
How does the NELL learning process work? Initially, it receives information that defines what should be its specific focus in such learning. “We inserted in NELL, in the form of input files, the concepts we are interested that it learns and what relationships between those concepts are important to us,” Hruschka explains. From there the program begins to ‘read’ files on the Internet to extract knowledge of specific subjects. To understand how this learning works, he explains how Nell learns the names of the world’s cities. “At the beginning we supply the computer with some reading tips that will help it identify cities in terms found on the Internet. We can tell it that every time it finds the sentence ‘X is a city located’, the word X refers to a city. After reading and identifying some cities, NELL is able to autonomously define new ways of identifying cities, using, for example, the sentence ‘The city hall in X’.”
In general, NELL learns facts that are a relationship between two categories, such as “I live in … (city, country, etc.).” or “he plays for … (volleyball team, soccer team).” The computer already dominates 280 types of relationship and that number is growing continuously. To avoid learning and propagating errors, all information learned is internally validated by means of a model of probability that considers the amount of evidence that that particular fact may be true and the amount of false evidence. It is this, for example, that helps NELL avoid confusing the name of a country with a city when it comes across the phrase “José Saramago was born in Portugal.” According to the authors of the research, an intelligent computer program like NELL could be used in countless applications. On the Internet itself, for example, it could give rise to more sophisticated search mechanisms that, instead of simply finding pages related to topics that we are looking for, provides answers to our questions. In companies, computer systems will be able to acquire experience and, just like older employees, accumulate knowledge that makes them more efficient over time. They can also be used as virtual personal assistants that learn the profile of their users and progressively serve them better – for example, playing the role of news assistants that automatically search the web for content of interest to the user.
The Read the Web project first appeared in 2008 during a visit that Estevam Junior paid to the laboratory of Professor Tom Mitchell, at Carnegie Mellon University. The two had known each other for three years. At the time, the Brazilian researcher was involved in a project on FAPESP’s Young Researcher program on databases. Although Hruschka’s project objectives were different, common research interests emerged and they started working in partnership. “In January 2008, when I arrived at Carnegie Mellon, Tom and I started the official Read the Web work. After the first year of work, we managed to define an initial architecture and the basic principles of the new endless learning paradigm. That’s how we began a NELL prototype,” says the Brazilian researcher. In February 2010 he returned to Brazil and started the Read the Web project in Portuguese, the results of which are likely to be integrated in the future with NELL. Here in Brazil the initiative is funded by the National Council for Scientific and Technological Development (CNPq). In the U.S., it receives funds from Google, the National Science Foundation (NSF) and the Defense Advanced Research Projects Agency (Darpa), the research office of the US Department of Defense. Yahoo gave them a M45 supercomputer and Microsoft Research is funding a PhD scholarship.
Other institutions and companies are funding programs with the aim of endowing computers with some kind of intelligence. Such is the case of IBM, which owns a project called Question Answering for investigating techniques that allow a machine to answer questions posed by humans. The team of Professor Oren Etzoni from the University of Washington, is also working on the extraction of knowledge from web pages, but with a fixed set of pages previously stored and without applying continuous learning techniques. “We have a great relationship with the two groups.” With regard to the functions of NELL, Hruschka emphasizes that “it does not have autonomy for any other type of action different from that of learning on the Web, storing this acquired knowledge and interacting with humans – or with the Web itself- to clarify doubts.” In other words, the risk that in the future NELL transforms itself into HAL 9000 or Skynet, the computer program from the movie The Terminator, which dominated the world by learning from the internet, is nil. “Changing how computers learn is neither simple nor fast. In the current version of NELL we already have a smart computer, but we believe that by mid-2014 we will be able to show a large part of the potential of the idea.”
Republish