Scientists plot AI that learns from mistakes

 

Gallery

By Stewart Mitchell

Posted on 20 Aug 2009 at 14:55

Scientists at Oregon State University are hoping to improve artificial intelligence with a project the uses "rich interaction" to teach machines when they make mistakes.

The researchers claim the project could lead to a computer that wants to "communicate with, learn from, and get to know you better as a person".

The software relies on computer users telling the machine when and why it has made an error in communication or logic so that it will automatically update its code in order to avoid making the same mistakes again.

"We want to develop algorithms that will allow the end user to ask the computer why it did something, read its response, and then explain why that was a mistake," says Weng-Keen Wong, a computer science professor on the project.

Ideally, the computer will consider the response and change its programming so that it performs better in the future. It's like debugging a program

"Ideally, the computer will consider the response and change its programming so that it performs better in the future. It's like debugging a program."

The researchers cite the way existing advanced learning systems customise themselves based on end-user input, but have no comprehension of why they are changing, which means they repeatedly make the same mistakes leading to false positives.

Such systems are the basis of spam filters, e-mail sorting and product recommendations such as: "If you liked this book, here's another one you might find interesting."

These intelligent systems are based on word statistics, set rules and similarities, the researchers say, but even the most advanced learning software only allows a user to tell the computer something is right or wrong, not to explain why.

According to Wong, the major stumbling block in the system is making the intelligence system not only interactive, but also easy enough for non programmers teach the computer.