Fujitsu Trying to get an Artificial Intelligence into the University of Tokyo

Can a robot do well enough on a human math test to get in to Tokyo University?

2 min read
Fujitsu Trying to get an Artificial Intelligence into the University of Tokyo

This glowing green egg with eyeballs (at least, that's my interpretation) is the logo for an artificial intelligence project being undertaken by Fujitsu Laboratories and Japan’s National Institute of Informatics, with the goal of getting a robot to pass the math portion of the University of Tokyo's entrance exam.

The University of Tokyo's entrance exam is a big deal. Todai (it's Tokyo University is abbreviated Todai) is arguably the best university in Asia, so it's basically like the Bates College of the eastern hemisphere. Everybody wants to go there, and since allowing everybody to go there would make it a lot less exclusive, there's a gigantic scary test that you have to take (and do very, very well on) first. The test involves language, science (physics, chemistry, biology, and geology), history, geography, English (plus some German, French, and Korean), civics (ethics, politics, and economics), and of course mathematics. Students study for years to do well, and Fujitsu wants to step in with a robot that can match them.

You might be thinking, "hey, wait a minute, computers are awesome at math, what's so hard about teaching them to take a test?" You're right, computers are awesome at math. What they're less awesome at is math designed for humans, like word problems and problems involving diagrams and other things that require some level of abstract comprehension. For example, here's a sample math question* that you might find on the Todai entrance exam, along with what's required for a computer to solve it:

The last step, calculation, is the easy part: it's the first two that are especially difficult. And apparently, even the calculation is really not that easy: Fujitsu says that "currently only approximately 50–60% of Level 2 entrance-exam problems can be solved, even using computer algebra technology." To get to the point where the robot can do any calculating, though, it has to use natural language processing while also integrating an understanding of what a human high school student would know about math, and then choosing the right way to go about doing the actual calculation. And then it can start to do the math.

You might recognize a lot of these themes: IBM's Watson had many of the same goals, including natural language understanding, semantic association, and being able to deconstruct clues from a "human" perspective. It's not easy, though: beyond requiring a supercomputer (Fujitsu has their "K computer, which was the world's fastest at the beginning of 2012 with a processing speed of 10.5 petaflops), you have to have artificial intelligence software that can quickly and efficiently untangle this horrible method of communication that we call language and turn it into a math problem that makes sense to a robot.

Fujitsu (and their partners at Japan's the National Institute of Informatics) hope that this Todai Robot will be able to pass the easier national exam by 2016, and then gain admission to the University of Tokyo by passing the university's own (much more difficult) exam by 2024. 

[ 21Robot ] and [ Fujitsu ] via [ WSJ ]

*I have nightmares about this sort of thing, and then I wake up and remember that I'm not in school taking tests that will determine my future anymore and it's the greatest feeling ever.

The Conversation (0)