The Todai Robot project made headlines a couple weeks ago for its goal of building an artificial intelligence system capable of passing the University of Tokyo’s stringent entrance exam, but developing the robot’s reasoning skills are only a part of the challenge. The team — which is headed by Japan’s National Institute for Informatics and includes IBM, Fujitsu Laboratories and others — is building the system to run on a laptop computer, which is just a fraction of the computing power that other famous AI systems such as IBM’s Watson and Deep Blue had at their disposal.
I spoke last week with Hirokazu Anai of Fujitsu Labs, who explained to me the thinking behind the Todai Robot project and the challenges in making it happen. At a high level, the project wants to create a system that can read and understand questions on the university’s entrance exams (Fujitsu is focused on math, while IBM is working on the history portion) and be able to answer them at a rate high enough to pass — currently between 80 percent and 90 percent for the first-stage exam, and then 30 percent to 40 percent for the second, more-difficult stage. The most-difficult problem, Anai said, is getting the robot to understand the questions and convert them into code that a computer can understand.
However, he added, trying to create a system that can carry out all the necessary calculations on a single consumer processor only makes the project more challenging. For the math questions Fujitsu is focused on, once the system converts natural language into computer-readable code, it then must choose from a collection of computationally complex computer-algebra algorithms the correct one to answer the question. If the system were running on a supercomputer, Anai said, the whole process would run a lot faster because of the increased computing power.
A workflow for a sample math question.
Through a translator, though, Anai said designing the exam-taking robot to run on a laptop means the project has greater utility once it makes its way into the public sphere. He noted manufacturing, insurance and even the travel industry as areas that could benefit by putting this type of decision-making power in the hands of regular users who might come across tough math problems in their day-to-day jobs. And even expanding the system to a small computing cluster could help generate answers significantly faster, still without requiring a major capital investment.
Feature image courtesy of Shutterstock user Palto.
