The Nature of Adaptive Robots

imtracynotstacy: Ok. First things first. How is your day going?
4:09 PM

j0shbongard: So far so good. It's been pretty quiet here today, so I've had a chance to get a fair bit of research done.

imtracynotstacy: Are you in the lab or in your office?

j0shbongard: In my office at the moment. My lab is in the building next door.
4:11 PM

imtracynotstacy: And what does it mean that you've gotten a fair bit of research done? What kind of research?

j0shbongard: There are several strands of work going on in my research group at the moment. The strand that I was working on today has to do with the idea of aesthetic preference: how people judge an image or object as more beautiful than another.

imtracynotstacy: And what does that have to do with robots?

j0shbongard: Good question. In our previous work with the Starfish robot at Cornell, together with Hod Lipson and Victor Zykov, we built the robot so that it could create models of itself. Those models were in essence small virtual worlds where it could mentally rehearse actions before attempting them in the real world.
4:13 PM
In some of our new work, we're creating online experiments in which a computer creates models of people's behavior. In the case of aesthetic preference, the computer learns from how a person responds to images, and tries to determine from that what criteria they might be using to judge beauty. [over to you]
4:15 PM

imtracynotstacy: Roger. So if a computer can accurately predict a people's behavior, for example, their aesthetic preference, how might that ultimately help the computer (robot?).

j0shbongard: The two projects are quite different. In the case of our robot, it can use its self-made model -- it's mind's eye, if you like -- to judge whether an action might be dangerous, and therefore discard it before actually attempting it in reality. In the case of our aesthetic preference experiments, the computer-made models of people's choices may reveal how they make decisions.
4:18 PM
In particular, it may help us to better understand how people make decisions based on criteria that are subconscious, or otherwise difficult to put into words. How one person selects one object as more beautiful than another is one example of such a hard-to-articulate decision.
[over to you]

imtracynotstacy: And where might such a computer-made models of people's choices be used? What are the applications?

j0shbongard: You might be able to bundle these ideas into software that can predict how people might respond in dangerous situations. Often, in retrospect, people find it hard to describe why they made a (possibly difficult) decision in a hazardous situation. [over to you]
4:22 PM

imtracynotstacy: Do I really a computer telling me not only that I screwed up, but how? :)

j0shbongard: Who knows, perhaps some day. :-)

imtracynotstacy: Great. Your work falls under the field of computer science, right?

j0shbongard: That's right: I'm a computer science professor at the University of Vermont.
4:24 PM

imtracynotstacy: How long have you been doing this kind of research?

j0shbongard: Since about 2000, when I was a PhD student in Rolf Pfeifer's lab at the University of Zurich, in Switzerland.

imtracynotstacy: What do you know now about this area of research that you didn't know in 2000?
4:26 PM

j0shbongard: I suppose what I've gained an appreciated for most is how little we understand about intelligence: whether we're talking about human intelligence or machine intelligence. [appreciated --> appreciation] Intelligence is not something that exists within our brains, but is something that arises from how we use our brains and bodies to survive, and excel, in an ever-changing world.
4:28 PM
It's exactly at the boundaries between the body and the brain, and the body and environment that our understanding of intelligence really breaks down. Our Starfish robot is a good example.
4:29 PM
It chooses its actions very carefully to learn more about its own body, and what that body is capable of. Those actions cause its body to move, which causes new sensor readings, which causes it to update its understanding of itself and the world around itself.

imtracynotstacy: So the more it acts, the more intelligent it becomes.

j0shbongard: Yes, exactly. But not only that: it needs to act in very particular ways in order to learn more. Just doing the same thing over and over again doesn't buy the robot much. For this reason, many people have commented that our robot appears curious.

imtracynotstacy: Interesting

j0shbongard: It's always trying to actions to learn more, or repeating a short series of actions to confirm a hypothesis. [to actions --> new actions]
4:31 PM

imtracynotstacy: It needs to take risks in order to grow. Just like a person.

j0shbongard: Well, not just like a person, but similar.

imtracynotstacy: Where do you get inspiration from?
4:32 PM