Interview conducted with Jerry Zhu, UW-Madison computer science professor.
Daily Cardinal: What's the state of the art in the field of machine learning?
Jerry Zhu: [Machine learning and artificial intelligence] is a thing you don't see a lot just from the surface. The [Defense Advanced Research Projects Agency] Grand Challenge with autonomous vehicles involves a lot of learning because they have to drive by themselves. Learning shows up in other things like speech recognition, when you get an automated voice. And a lot of things happening behind popular search engines, like Yahoo! and Google, actually involve these [learning] techniques.
We're trying to make the machine be able to learn as a human does. Right now, you have to teach the computer a lot. You have to hold its hand and tell it what to do.
Think of driving. The computer sees an image, and you have to teach it how to turn the steering wheel. So it learns that if it sees the road curves to that side it's supposed to turn the wheel this way. But you as a human teacher, you have to do that operation.
What I'm trying to do is reduce the amount of teaching to the minimum possible amount. Maybe only teach it for a minute [instead of an hour]. And then let it learn itself.
DC: What are the trends?
JZ: [Like any research these days,] I would separate trends into two components. One [major component] is to make what we do now a little bit better, and the other component is to do things we cannot do now.
You see a lot of work done jointly. There is a trend that the field that studies artificial intelligence—that's computer scientists—and the community that studies human intelligence—neuroscientists and psychologists—are doing research together.
DC: What sort of problems does computer science face?
JZ: We do not have a true, deep understanding of human intelligence; we do not know for sure how we think and make decisions. Artificial intelligence is mostly based on mathematics and logic. These are certainly related to intelligence, but we still do not know what true intelligence is and how to duplicate it.
A different issue is the expectation of the general population. There can be a huge mismatch between what a computer can do now and our expectations, defined by popular movies, of what computers should be able to do.
We tend to think a computer is either very smart—that it can drive the car for you and robots that can talk to you—or to the other extreme, that computers are just too dumb to do anything correctly.
We need to understand there is only a limited set of things a computer can do well—and quite well, but it's hard for them to do it outside of this set.
