Push Singh is interested in the hardest challenge of Artificial Intelligence, common sense. While programming a computer to beat humans at chess or index a million documents are solved problems, it's still remarkably difficult to teach a computer than rain is wet and that humans don't like to be wet. Speaking today at the PUSH 2005 conference in Minneapolis, Singh is showing off the work he and AI pioneer have been recently engaged in.
At MIT's Media Lab, Singh launched a project called OpenMind, which is a collaborative, participatory common sense knowledgebase for artificial intelligence. OpenMind now contains more than 750,000 "items" of knowledge - factual assertions - contributed by roughly 16,000 users. Users register to use the system and are presented with tasks that are comparatively simple for humans - writing a two-line description of a photo, adding five "facts" to explain a statement like "the boy drank the milk".
The OpenMind system takes these user-contributed facts and uses them to build a set of tools - ConceptNet, LifeNet and StoryNet, each of which is an interconnected set of ideas that can help a computer solve common sense problems. ConceptNet links concepts together to help a computer understand what a natural language term means; StoryNet and LifeNet link situations together to let a program understand what events might lead to other events. Other tools include GlueNet, which help identify phrases that have similar meanings, and ShapeNet, a distributed approach to tuning computer vision.
OpenMind isn't the only system attempting to solve problems like this using the agglomeration of common sense facts. Doug Lenat's Cyc database has collected hundreds of thousands of "assertions" since the early 1980s; OpenCyc now is attempting to replicate some of this work in an open source environment. Singh describes OpenMind as being different in that it's taking a wiki-like approach, noting that it was accepting user contributions before Wikipedia, in 2000.
(Push and I briefly talked about quality problems with user contributed material last night. He's been pleasantly surprised at how little bogus data has been fed into the system. I'm wondering if this is because, unlike with Wikipedia, it's very hard for anyone to see grafitti posted on the site. Your input goes into OpenMinds, but it's hard to see where your input goes and what it does...)
As the database gets richer, it could be useful for building intelligent search engines, cameras that know when they should take a photo, or PDAs that know to cancel appointments when you're out of town. I wish that Singh had a bit more time to show us some of the applications he and the Lab are using for the OpenMind database - it's clear that this is a really interesting technique for assembling a body of common sense knowledge and less clear what this knowledge can eventually be used for.