Wednesday, May 21, 2008

Artificial Intelligence

The 20- or 30-year-ago promised advances in Artificial Intelligence have not been realized. We have no robots, no near-human strong AI, no autonomous vehicles, no universal voice recognition and universal translation. We still must clean our own houses, guide our own vehicles, perform our own surgery. The good news is that we still have jobs.

Certainly we have made progress. Voice recognition is gaining traction, although the vocabularies are severely limited. Autonomous vehicles have been demonstrated (see the DARPA Grand Challenge) on a very small scale. Likewise, automatic translators are now useful although seriously flawed. Chess playing machines can beat any human, although through brute force, not insight.

But we have no idea of how to design a self-aware machine, or a generic learning machine, or a replicating machine. We have major difficulties duplicating simple human tasks such as identifying and understanding a speaker in a noisy environment (as at a party). We have had nearly zero success in recognizing objects in a jumble, such as a specific toy in a toy box. And don't even get me started on tasks such as bipedal walking or running, special effects movies aside. Lip reading or sign language? Hah!

There is hope. The recent improvements in voice recognition have come from brute force, throwing huge computing resources at the problem. As computers continue to improve, we can expect similar progress in visual and aural recognition, autonomous motion, even humanoid robotics.

Some Expert Systems (Inference Engines) have shown impressive performance in limited domains.

Some large Neural Networks have resulted in valuable insights, including speech recognition. But these are learning systems. We teach, but we don't truly understand.

Thinking, consciousness, creativity? We don't know how humans do it, let alone how to teach it to a machine.

Progress will be frustratingly slow because we simply do not know how to do a great many tasks that humans find intuitive and trivial. But perhaps in thirty years when we can throw a million times the CPU performance and a million times today's memory capacity at this problem, our computers will figure it out on their own.

In a sense, we will then have created strong Artificial Intelligence, but we won't understand or control it. Sorry.

Many people have written stories about creating truly intelligent machines. I'll soon be one of them: when my story The Awakening comes out, please read it.

No comments: