Monday, March 16, 2009

A.I. In The Enterprise

Artificial intelligence will likely find its way into corporations.


Smart machines and robots that can think and react more quickly than people have been the stuff of science fiction for decades. But only recently has there been enough computing power and memory available at a reasonable price to allow it to progress to the next step.

Tests are now underway in the military to create a smart surveillance system that can interpret facial features to identify people, determine what movements are unusual and sound the alarm where necessary. And that's just the beginning. The technology will find its way into government, corporations, and eventually, even the home.

In business, these tools will almost certainly will fall under the domain of CIOs as part of their expanding role in enterprise information management. But this kind of information is somewhat different than what CIOs have dealt with in the past. It still uses computers, databases and data mining, but the method of gathering information and its application head in a sharply different direction.

Forbes caught up with Rachel Goshorn, assistant professor of system engineering at the Naval Postgraduate School in Monterey, Calif., to talk about artificial intelligence and what's changing.

Forbes: Why has artificial intelligence taken so long to get out of the labs and into the real world?

Goshorn: It has been used on a limited scale for inspecting food and things like solder joints, where the rules were simple, but the computing was so intense that, for a long time, it didn't get much further.

Now that computers are cheap, it can be applied for a lot more markets, right?

Yes. One of the things that changed is that before, there was never the ability to make a sequential behavior recognition model. It was static. There weren't the programming techniques to build a model. Everything worked like a flow chart. It gave you the option of "yes" or "no." The kind of logic we needed from computer science didn't exist. We had to create a matrix so you could keep adding to it and give it statistical weighting. We needed the ability to detect, identify features and then predict behavior and react.

All of which are incredibly compute-intensive, right?

Correct. AI is not a single thing. It's algorithms for detection, identification, prediction and reaction. These are all working in parallel. That's why you couldn't do it with a flow chart.


How does this work?

It's a sequential behavior classification system. We started out by classifying cars on a freeway. At time one, what lane are they in? At time two, what lane are they in? You concatenate where a car is over time, and you get a pattern. That allows you to predict possible accidents based upon patterns. The behavior classifier fuses electrical engineering, computer vision and compilers. You have to design an alphabet, and each letter has to be assigned to a feature or a symbol with time and space associated with it. Behavior is how you put these symbols together. It could be all A's if you're in the slow lane. A behavior is an infinite set of sentences all made up of these symbols.

Then do you match this against so-called normal behavior?

Yes. Behavior is not 100% predictable, so you have to match it against a measure of something known. If it's too far away from something that's known, it can be flagged as abnormal. This modeling has to be problem-dependent. You use the same AI system-level algorithms, but they change depending upon what application you're using it for. The first phase is you get some raw data, maybe from sensors, and you have to detect certain features.

So this is all about probability rather than definitive answers?

Exactly.

What are the real-world applications of this?

In the radio frequency world, we were able to detect when a frequency was a friend or a foe. If you're in the middle of the desert and you've got the frequency of a garage door opener that's used to set off an IED [improvised explosive device]--you know historically these are used to initiate IEDs--you can point to where that frequency is coming from, and you can jam the frequency so it will not ignite the IED. You also can send a missile to that spot.

In the past, much of the technology that ended up in the commercial world came out of the military. Where do you think that will happen with artificial intelligence?

In the industrial world, you can do it for quality inspection. It also goes into marketing and analyzing. Inside large warehouses or grocery stores, you can learn the behavior of shoppers so you can place products in different places in the store. You also can learn the behavior of online shoppers to reach them more effectively.

How does this work with existing equipment?

If you have a microphone and a camera, they don't have a lot in common. To use them as inputs, you have to do a lot of programming. But if on the microphone you could put detection, you don't send back the voice. You can identify where a person is so the camera zooms in on him or her. They can be correlated. The idea of bringing multiple sensors--they could be voice, image, temperature or pressure--and extracting features rather than pixels or analog voices is all doable. We do this in our brains. Once you translate it into something that isn't analog or a pixel, you're into the human system. It smells like mustard or it's a brown dog. Once we classify it, we're through with the sensors and into the realm of intelligence.

What's the short-term and long-term use?

The immediate use is Homeland Security. Sensors are becoming so cheap. If you're in Southwest Asia and troops have to go over a hill, you want to make sure it's safe. You can drop cheap sensors--microphones, cameras, magnetometers. Then you pull out the features to know what you're hearing and seeing, and over time and space those features make up various behaviors. Is everything OK over the hill? Are there people there? Are they good guys or bad guys? You can do the same with the border patrol here in the United States. Are they carrying bombs?

You can take any market, whether it's quality control, manufacturing, financial or medical, and you can apply it. That's what managers and executives do. If you're buying something in Phoenix and San Diego at the same time and you've never been to either of those cities, an alarm comes up. American Express applied AI years ago. In the financial world, if you want to model all the variables of shareholder selling, inside selling and everything with buy and sell, you can start predicting behavior.

How about the consumer world?

You can see an AI babysitter, dog-sitter or house-sitter. Based on certain behaviors, you can set off alarms and react. You can measure temperature and pressure better than intensive care in a hospital. This works for the elderly as well as babies. You also can do this on the Internet. A lot of videogames are using this for multiplayer modes like fighting.

Ambient assisted living is another area. You can enable smart rooms and look for the behavior of the elderly. If someone is bedridden at home and wanted to communicate with hand gestures to turn on the TV or call the doctor, hand positions could mean different things and a sequence of hand positions could mean different things.

Does this make robots more viable?

That's where AI was first applied. They were expert systems. The whole industrial automation world was created to build smart robots.

No comments:

Post a Comment