Modelling minds

Toby Howard

Chris McKinstry is a man with a rather unusual mission. He wants to use the Web to create a true artificial intelligence -- a working model of human thought.

For his day-job, McKinstry runs the European Southern Observatory's Very Large Telescope in Chile, but it's his Mindpixel Digital Mind Modelling Project which is attracting attention.

McKinstry is a veteran of the annual Loebner Prize competition (www.loebner.net), in which artificial intelligence programs compete for the best performance in a Turing Test -- attempting to give responses to questions which would be indistinguishable from a human's. After becoming disillusioned with the Turing Test, which he believed was testing human-like interaction rather than basic human reasoning, McKinstry decided on a new approach: to make a model of the mind based on simple facts which are either true or false. McKinstry calls these facts "MindPixels".

McKinstry believes that what makes us human is our knowledge of the world around us. In his view, what we regard as "truths" about the world, are built up from layers upon layers of more simple truths that ultimately hinge on facts which are either true or false. So, he reasons, if you can create a database of knowledge which contains a huge number of yes/no facts, such a database will effectively encapsulate a view of the world. This is what the MindPixel project is trying to do.

The first phase is to collect the MindPixels, and that's where the Web comes in. McKinstry wants people to visit his Website (www.mindpixel.com) and talk to a programme running there called GAC, for Generic Artificial Consciousness. When you "talk" to GAC (pronounced "Jack"), you submit a MindPixel of your choice -- "Personal Computer World is a jolly good read", for example. Subsequently, this MindPixel will be presented to other people for validatation. And in return for submitting your own MindPixel, you're asked to validate a set of twenty others, stating whether each is true or false. To give you an idea of the kind of facts GAC is learning, here's a few MindPixels I validated recently:

Drinking ammonia is good for your health.
Is today Monday?
Is Istanbul a city in Turkey?
Oporto is in Portugal.
Does a mink have fins?
We all need love.
Milk does a body good.
Is it always dark at night?
People are larger than ants.
Panama is an isthmus.
Does a deer have feathers?

The MindPixel project is distributed computing of a kind: "It's like SETI@home in many respects," says McKinstry, "but we're not after your CPU's cycles, we want your your humanness". And it's certainly catching on: 16,500 people have registered, and GAC's database already contains over 136,000 MindPixels.

Once a big enough database has been built, and McKinstry estimates he'll need a billion validated MindPixels, the second phase of the project will begin, probably around 2010. The database will be used to train a neural network -- and the end result, says McKinstry, will be a model of the human mind that can respond correctly to queries which are not directly answerable by the MindPixels in its database.

McKinstry's isn't the first attempt to create a database of human common-sense knowledge. Cycorp's Cyc project started in 1984, with similar goals. But the Cyc database, which currently contains 400,000 cross-referenced facts, is being built entirely by hand, by a small number of "knowledge engineers". Cyc learns "top-down", as its programmers encode facts in carefully-designed interrelated tree structures. In contrast, GAC learns about the world in a chaotic "bottom-up" manner.

What GAC will eventually turn out to be like, is anyone's guess. But even if it's never truly intelligent, it'll be very handy to have around for a Pub Quiz.

Toby Howard teaches at the University of Manchester.