Artificial intelligence gets ready to party

A San Diego company has created AI software that can assign gender to faces and pick voices out of a noisy babble. And it could have serious applications too
Written by Rupert Goodwins, Contributor

Cocktail parties and relating to the opposite sex are two very human activities that preoccupy many during the holiday season. Computers are notoriously poor at both: artificial intelligence researchers have long envied the human abilities to assign gender to faces and pick voices out of a noisy babble.

Now San Diego company HNC Software has taught personal computers equivalent feats of recognition, and claims that the technique is a significant advance for AI, both theoretically and for building practical applications.

The system, called Cortronics, uses neural networks in a ring of Pentium-based PCs to correlate inputs with memories of previous inputs, checking for coincidences in both time and space. Items the network perceives -- which can be words, images, financial information or anything that can be expressed digitally -- are broken down into tokens that are stored in a database. Tokens that coincide in time or space are associated, and gradually the network builds up a hierarchy of how these associations themselves are linked together. When objects always appear together, the system recognises that the group is itself an object, and over time can build up subcategories of that object such as male and female faces. It uses this information to recognise new occurrences of objects that contain a similar collection of tokens, an essential part of intelligent behaviour.

The system notices when some tokens in an association are missing, and pays close attention by watching out for the others, which improves performance. This means that recognition is enhanced when objects are partially obscured, for example. Similarly, when the system recognises some words in a party babble, it works out what words are likely to follow. If one of those words follows, it knows that it's tuned into one specific speaker and then concentrates on that.

The company says that these behaviours are close to many of the ways that humans perceive real-world data, and that the technology has great potential for image recognition, automatic customer conversational services, security and so on. The product isn't a turnkey solution, however: companies wishing to use it must send engineers on a three-week, $50,000 training course with HNC to learn how it works and how to integrate it.

See the Software News Section for the latest headlines on everything from peer to peer clients to Office software and beyond.

Have your say instantly, and see what others have said. Click on the TalkBack button and go to the Telecoms forum.

Let the editors know what you think in the Mailroom. And read other letters.

Editorial standards