Fabio Cozman

Fabio Cozman

As part of our AI in Your Community project, I spoke to Fabio Gagliardi Cozman, who is a Full Professor at the Engineering School  at the University of Sao Paulo, Brazil. He  works in the Department of Mechatronics and Mechanical Systems, in the Decision Making Lab, which focuses on Artificial intelligence.

Tara Chklovski: Tell us a little bit about what kind of problem you’re solving.

Fabio Cozman:  I’ve been working with artificial intelligence for about 25 years now. I did my PhD in the United States at Carnegie Mellon University, and back then I was working with robots on a NASA-funded project where the idea was to send robots to the moon or Mars to look for rocks. I worked with the cameras to detect objects of interest or potentially dangerous objects.

Then I came back to Brazil and now I work with machine learning. Machine learning is basically getting lots of data and extracting patterns from data. More recently I’ve been working on combining data with knowledge a person might have. For instance you may have a doctor that has some information about a disease and you may also have data about patients with that disease, so figuring out how to combine those two sources of information is the driving force of my work.

I’ve been working mostly with large databases, like the Never-Ending Language Learning or NELL database. Nell is a computer system that learns over time how to read the web. It’s based at Carnegie Mellon but there are people all over the world working on it, and there are many of these sorts of database around the world, containing millions of facts. These are the common knowledge of humanity, so to speak.

For another example, let’s take a project I worked on a few years ago which helped students find scholarships. One one hand you have  system that has to interact with the public to help students find scholarships, so the system has data to reason with, but on the other hand you also have knowledge about how humans think about problems. So, combining those things effectively is the key.

TC: Can you give an example for that scholarship project? What types of common-sense facts would be helpful in that case?

FC: For example if a person calls and says “I’m looking for information” you’d have to understand that the person’s looking for information about scholarships. If the person says something like “I have two kids and I’m interested…” you know she has two kids, and that she’s looking for information for the kids, not for herself.  Sometimes we would get calls from the students, sometimes from parents, sometimes from grandparents, sometimes from professors, teachers. So, some of the information you get from those calls gives you a lot of things to  reason out. You have to be able to do that.

TC: Right. And humans do that automatically.

FC: Yes. It’s easy for us.

TC: So a big challenge that we have seen when we run our programs is that children get stuck on the very first step of finding a problem that they can solve. What advice would you give children when they’re trying to find a good problem to work on?

FC: I’ve thought about this a bit: what could be a good problem? Because when I try to find a good problem to solve, usually I take something I need to be done and I try to do it. When you do it, and you try the tools that are available, usually you find that they are not as good as advertised, and then you realize, wow, this piece is missing.

If you have a  kid that is interested in writing a little code to, say, help the school decide which texts are preferred by students or maybe in writing a little code to help teachers decide which math problems are more challenging, I think I would, as a kid, try to put myself in the shoes of the teacher and think “well, what information could I use?” When you try to solve the problem by yourself, then you realize, oh, this is missing, and that’s missing.

TC: And so in your opinion what are some elements or aspects of a good product?

FC: That’s tough. I will say a few words about a good AI product. I think one problem that we have with AI is that sometimes it’s obnoxious, like when you have assistants that help too much. They try to help you put the email in the box or decide what to buy. I think a good AI product is one that does solve a problem, not something that is just there to make you feel amazed for a while. It solves a problem without bothering you. I think you want something that helps you, but it shouldn’t be obnoxious.

TC: We’ve been working with student groups that have been coming up problems and AI systems to help, and they quickly realize they need to understand human psychology. How does a computer scientist begin to develop an understanding of human psychology, because it’s a completely different field?

FC: I think this is very important today, as we have more and more systems that attain human ability. For example, now we have systems that are better than humans at recognizing faces, or deciding which person is in a picture. So, as we develop such powerful systems it’s very important that their decisions can be explained. That interest in how humans explain and interpret things is something that we should find in psychology. There is actually quite a bit of interest in this question: how is it that we explain? How is it that we feel satisfied of an explanation, when we feel satisfied?

TC: Do you have any recommendations on how children can start learning about psychology?

FC: First of all, one nice way is to go out and play with LEGO Mindstorms and other robots where you can actually program a little bit. Most of the tools in AI are related to computer programming, so I think it’s essential that you understand what it means to program.

And then there are some systems, talking about machine learning specifically, that are geared towards teaching. For instance, there is a package called Orange, from Slovenia.

TC: Thank you! That’s a good recommendation. In your world what have been some major difficulties that you’ve encountered and how did you overcome them?

FC: What I find really difficult is finding a research direction, which means finding a large problem that would help people if it was solved, but also something that is solvable because it’s easy to find things that nobody can solve. And finding this sort of research direction is difficult. It’s frustrating.

When you have a research direction then I think it’s not hard to find problems; it’s not too hard to find solutions. I mean of course, you have to work; of course, sometimes things go badly, but that’s not as bad as when you don’t have a research direction. I would recommend  just thinking about your plans. You have to plan, you have to think about what you want to achieve in the future. There’s no substitute for reflection and planning in that regard.

TC: How long do you typically work on one research direction?

FC: I tend to stay with problems for a while – maybe five years before I change course.  

TC: And so what inspires you to keep going?

FC: Well I like research, first of all. I really like it. Speaking specifically of AI, I’m an engineer, I like getting things done and built. But I think with AI what attracts me is the interest in understanding intelligence –not so much understanding human intelligence, but learning about what it means to make an intelligent decision, and understanding enough about it so that you can reproduce it.

I think that’s the beauty of AI. It’s very nice in that sense. You get to address problems that have been with us for thousands of years with a different perspective.

TC: So just two last questions. What new field of AI excites you a lot?

FC: Well now I have to say this: it’s the combination of knowledge and data –machine learning, but I would say knowledge-based machine learning, which is I think very challenging.

I have a few projects I’m working on related to this. There is a text-based understanding project, which is more along the lines of the one that was helping students find scholarships. And I have one about social networks where we’re trying to decide who should be connected or who could be connected and then try to explain why.

TC: Last question – what advice would you give children and parents who are going through this challenge that could help them persevere?

FC: This is very hard – I have a daughter who is in school and I face that problem myself. I think it’s difficult for kids in Brazil, especially girls, to find themselves interested in computers because all of the boys are doing computer games and just shooting people in computer games.

I think what would be nice for parents of daughters would be to have them interact with computers but not in the context of computer games, just try to make computers more interesting by showing the girls interesting programs to work with, or things that are useful that they may use. I realize this is a big challenge.

TC: Thank you so much.