Many AI researchers believe that the biggest barrier to creating human-like intelligence is that humans know millions of simple everyday facts. This ordinary knowledge ranges from knowing what a horse looks like to a simple fact like “people buy food in restaurants.” In the past, AI researchers would spend years painstakingly entering such information into huge databases, but now a new crop of researchers are leveraging the millions of Netizens who have nothing better to do than answer stupid questions all day to build these databases quickly and for free. One such site is the OpenMind Initiative (hosted by my own Ricoh Innovations), which is primarily being used by the MIT Media Lab to collect Common Sense Knowledge.
The latest foray into this space is the ESP Game. When you log into the game you are paired randomly with another player on the Net. Both you and your partner are shown the same 15 random images from the Web, one at a time. Your job is to type in as many words to describe the image as possible, with the goal of matching a word your partner has entered. When you agree on a word, you both get points and move on to the next image. Usually I don’t care for Web-based games, but I have to admit the game is compelling.
The real goal of the system is to generate a huge database of human-quality keywords for all the images on the Net. The task is huge: Google’s Image Search has already indexed over 425 Million images by using the text that surrounds the image’s hyperlink. But numbers are on Ahn’s side: if only 5000 people were to play the game throughout the day, all 425 Million images would receive at least one label in a single month. Given that many game sites get over 10,000 players in a day, a few months is probably all Ahn needs to fill out the whole database.