I expected science fiction from Eli Abir, not fast food. Abir, who is not a well known scientist or technologist, says he has solved a problem that has caused quite a number of artificial intelligence scientists to tear out their hair. He has invented a “translating machine” – the first artificially intelligent computer brain capable of translating texts from one language to another on a high level. So where does the fast food come in? Abir says, “Now I’m going to set up a fast food chain for baking bourekas pastries in New York.
”I’m going attach sensors to the employees’ every muscle, which will listen to their hearts and monitor their responses. The computer brain will sit on the side, watch them manage the store, and learn. The idea is that the brain will later manage the bourekas shop by itself. This will be its first job as a robot.”
”Globes”: You’re going to set up a bourekas chain just to teach this brain how to manage this kind of business?
Abir: ”The reason I’m starting with bourekas is that it’s a fairly simple process, with relatively few things to do. Before I teach it to be a pilot or a doctor, I’ll teach it to do this.”
”It happened by accident, and scared me to death”
Even though the translation machine is still under construction and its current results are limited, it is arousing interest in the US scientific community.
In order to put the method to the test, Abir contacted Language Technologies Institute at Carnegie Mellon University director Prof. Jaime G. Carbonell, one of the leading artificial intelligence scientists in the US. Abir says Carbonell was very enthusiastic when the method was presented to him, and said he had been looking for years for the missing link that would advance artificial intelligence research to the next stage. Up until now, the various methods have achieved at most 80% accuracy; they have not purported to accurately imitate the way the human mind works. Carbonell recently decided to abandon his academic research and join Abir’s development efforts.
The odd thing is that such an ambitious invention came about quite by accident. Abir says it began as an idea for registering Internet domain names in various languages, on which his company Internet Driver was based.
”The way things happened with us was very out of the ordinary,” Abir says. “Usually, you begin a project, concentrate on it, and don’t leave it. The first patent I developed was an add-on to the Internet browser that enables people to write URL addresses in Hebrew. The system we developed analyzes and breaks down the URL address into all of its components and translates it into various languages. Later, when we began looking for money to complete the system’s development, people asked me why I didn’t use it to translate content. That’s how I actually began thinking about the brain.”
Abir says it took several months before actual development of the systems began. “This is something that actually works like a human brain, and that scared me to death, so for a long time, I didn’t want to do it. When I told people about it, at first they thought I was crazy, but then they convinced me to do it. I sat down and wrote it. It was so revolutionary that we decided to develop it at the same time as the other things. I wrote the patents in March of last year.”
Abir set up Fluent Machines, a subsidiary of Meaningful Machines, and a fellow subsidiary of Internet Driver. He later set up Future Applications, another subsidiary, which develops applications based on the artificial intelligence technology. The company’s financing rounds to date include a $300,000 pre-seed investment from a private Israeli investor. In the subsequent seed round, US investment company Apple Core invested $2 million at a company value of $9 million, before money. Apple Core is a very well known Internet investment concern, whose previous investments include companies like Register.com.
This financing round was later expanded with the help of another $500,000 from Apple Core and another private Israeli investor, at a company value of $25 million. Apple Core invested a further $2.4 million in the next round at a $40 million company value.
Meaningful Machines is currently holding a financing round. Abir hopes the sum raised will be used for constructing various applications on the basis of his translation machine. As one example, Abir offers, “It was recently reported that the US government has been sitting on untranslated Arabic texts since the first attempt to attack the World Trade Center in 1993. They haven’t had enough people to translate all the material. Our system would have solved their problem.”
Discovering the DNA of a language
The system’s technologies are currently in the final patent approval processes in the US. Abir’s translation machine searches for correspondences between word strings in the original text and strings in the translated text. The process assumes that when a word string is repeated several times in a given text, it will also repeat itself several times when the text is translated into another language.
The next step is to break down the word strings in both languages suspected of corresponding to each other into smaller units, down to the level of word pairs. A further search for correspondence is then implemented. In this way, every word pair undergoing the process is considered a “verified idea unit”. The next word pair, which consists of the last word in the previous pair and the next word appearing after it, must fulfill the correspondence with the previous pair. One example of this method’s uses is that it can accurately translate English verbs into Hebrew in either masculine or feminine conjugation.
At the end of this process, which involves several other innovative methods of textual analysis, what we get is a kind of language DNA. “The system actually locates the idea kernel beneath the language,” Abir explains. “If I want to translate a certain sentence from Hebrew to English, and I don’t find a translation in the existing database, the system will study it in French, Russian, and several other languages, then translate it back into English. Actually, the more languages the system translates, the more it learns.”
It can also learn to make coffee
The distance between an innovative translation machine and true artificial intelligence in the brain of a robot managing a bourekas store without human assistance is still very wide. Typically for him, Abir doesn’t really see the difference. He says that in the same way the system links sentences in two different languages to each other, it will be able to link physiological parameters with the actions of the employees in the bourekas shop.
”One of the things that will happen in this experiment is that everything the employees do will be recorded – every movement of every muscle, every emotional response to what happens in the store, every verbal response, and so forth. There will also be someone sitting and recording each process by hand. For example, the brain will recognize that if someone asks for a bourekas, you don’t take a chair and throw it at him. When the brain sees an employee move quickly, it will realize that it’s because someone shouted at him. The brain will slowly identify common denominators, and create a generic behavior, which will be the average of all behavior viewed by everyone observed. If the brain observes 100 people making coffee, it will study all the methods used and choose the way most people do it as the best.”
But the brain is only learning how to manage a bourekas shop. That’s still not real artificial intelligence that can take things from one area and use them in another area not previously studied.
”I’ll say someone a bit futuristic now – when I build one smart application, followed by another and another, at some stage the brain will be able to scan them all, which will give it the ability to derive things it has never done before.
”I don’t want to frighten you, but when I invented this, I didn’t sleep for a week. I was afraid for one simple reason – I don’t know what will happen to humanity, because with time, the brain will be able to do everything. Many things will change as a result. A person’s need to get up in the morning and do things will vanish because of this thing, and that’s frightening.”
How exactly do you plan to do it? Tell me, for example, how the brain can learn to play chess by itself and defeat IBM’s Deep Blue supercomputer?
”No computer, even Deep Blue, understands three dimensions. I don’t know exactly what algorithm it uses or how it thinks, but I know that for me to build a computer that will beat everyone in chess is the simplest thing in the world. It will simply take the information from other sources. It doesn’t even have to see a chess game; it only needs to learn the basic rules.”
Buy you have to feed it such a huge amount of information that it becomes impractical.
”It will know everything about everything. You’re talking about by far the most powerful computer in the world. We’re talking about something that will become the largest database in history.”
”For my inventions, prior knowledge is just an obstacle”
Abir’s faith in the road that will enable him to realize this ambitious goal probably derives from the fact that he is completely unaffected by what he describes as “harmful influences”. He says that a person in the computer field who enters the field of artificial intelligence will be unable to cast off the previous knowledge he has accumulated, and will attempt to achieve the goal using the methods he has employed in the past. Abir says this road is doomed to fail.
Abir is a typical self-educated man. He says, “I was the worst history student in the history of Israel. I have 11 years of formal education and even less practical experience.” He left Israel for the US at age 23, founded a prosperous used car business, and later set up a restaurant chain, which still exists.
Internet Driver was Abir’s first technology venture, which later led to his ideas in artificial intelligence. Without any formal technological education, Abir is now applying the method he developed in the field of computer translation. He doesn’t regard his lack of education as a handicap and says, “I turned a computer on for first time in 1993, when my son entered first grade”.
How does someone like you, without any background in computers or algorithms, direct the team of programmers that is supposed to make this thing work?
”I don’t write algorithms. I break everything down into stages for them, into the smallest steps. I control the whole process. The algorithm professionals are the least capable of doing this. Sometimes I give them a larger picture, and they don’t understand it, so I cut it down until I know they do understand it. I work with a lot of people that way. They develop nothing except for the next five minutes, until I see that it works and explain what comes after it.
”A lot of professionals thought at first that I was crazy. One expert, who now works for me, was sent to me by an outside concern that wanted to examine the technology. He had never seen anything that works like this. He told me he expected to see an entire book of software. The algorithm for this whole system fills maybe five pages.
”It’s a significant breakthrough, but it still needs to be tested.”
Does Carbonell, who believes in Abir’s project and recently even joined it, regard Fluent Machines’ technology as a significant breakthrough in computer translation? “Research has been going on in the field for over 50 years, and it’s not easy to discover new breakthroughs like this” he says. “Abir’s development in example-based machine translation is very promising.”
How does it succeed where other technologies and research theories fail?
Carbonell: ”It’s not that the other technologies fail. Labeling machine translation as a success or a failure is a very simplistic attitude. Today, for example, you have interactive translation machines that rely on earlier technologies, like the ones you can see on Google and Alta Vista. These technologies were successful to a degree, translation machines still need a higher degree of accuracy for general purposes.
”The time and effort required to build a computer translation system for two new languages is very great. Abir’s system is designed to significantly improve two aspects: the level of accuracy and the degree of effort required for development. That’s what makes it a potentially very important breakthrough. That is a really great achievement.”
Can you estimate the potential accuracy this method is likely to achieve?
”It’s possible. The US government National Institute of Standards and Technology (NIST) periodically gives estimates for new machine translation (MT) technologies. We are currently discussing how and when to participate in one of these estimates.”
Can this technology be used soon for practical uses, or is it still a theoretical promise?
”I believe practical applications can be achieved soon, assuming there is sufficient financing for the development period and Abir continues to invest his inexhaustible energy on technical guidance for the project.”
Carbonell’s report, which reviews Abir technologies, also covers the various methods used up until now in computer translation. These methods include transfer-based MT, statistical MT, and example-based MT. Carbonell classifies Abir’s system, which he calls EliMT, in the third category, and details several of its significant innovations. He signed the report with a caveat that Abir’s technology must still be tested. At the same time, he described it as “the most important theoretical development in the field in recent years, probably since computer translation was invented.”
Published by Israel's Business Arena on April 4, 2002