Part eight of our AI Researcher Profile Series invites Professor Klaus Mueller of the Department of Computer Science to discuss his extensive work with Artificial Intelligence, and how he uses it to produce a human-machine relationship that enables the machine, but empowers the human.
AI Institute: What sparked your interest in Artificial Intelligence?
Professor Mueller: You can’t escape it because it’s everywhere. I’m actually an electrical engineer by training, and then got into computer graphics and medical visualization in graduate school. I then started to do some information visualization and visual analytics. There I realized that humans can be supported by machines. Humans and machines can be partners. I don’t want to eliminate the human, but rather enhance the human’s abilities to enable human-machine teaming.
Once you do this machine assisted human reasoning, there’s a lot of things you come across, like bias in the data. People give too many tasks to machines. You cannot trust the machines because they’re trained with data which themselves are imperfect. Causal and deep neural models are imperfect because the data they are trained with are imperfect, but there’s just not enough data around. They come from humans who are biased themselves. We try to come up with frameworks that empower the human, but still keep the human in the loop so the human is basically the one who is really in charge and keeps the models on track.
AI: A lot of your recent research involves visual analytics. Can you share some new and exciting developments you’ve seen?
PM: Visual Analytics is about giving tools to experts to explore the data that they have. One exciting thing was when I worked at Brookhaven National Lab as an adjunct senior scientist. I gave a talk at the l;ab where I met two atmospheric scientists. They had this machine that could suck in a lot of atmospheric particles, millions of them. They would do a spectral analysis of these readings and in the process created a very high-dimensional data vector for each particle. They could categorize chemical elements and molecules from that vector. They would take a plane, go into the air, and measure what particles are out there to figure out what causes the CO2 to be absorbed in this atmospheric level, and then figure out why the sun’s radiation gets reflected a lot. They often told me that the reasons for the greenhouse effect are much more complex than people think.
We built them a visual analytics system that would let them tune the machine learning models they used so they could actually work with the noisy data they had. They got several grants from the government, and their data-driven work became a real use case for visual analytics. We published several papers with them. Without visual analytics, they couldn't have done what they achieved. There are many more examples of how visual analytics can help- it’s basically coping with the fact that automated systems are not powerful enough because the data are often incomplete and noisy. The human has a lot more knowledge about the process at hand. Let the human be in the loop of that.
AI: Over the years, have you seen a shift in topic interests among your students?
PM: There’s more focus now on people that want to do Human-Computer Interaction (HCI). I have quite a few students that are interested in this—understanding how the human operates, and making systems that have humans in them. Social good is also a topic that has come up a lot. Several students want to work on computational fairness. They’re really into it. They can’t see themselves doing anything else.
People also like reinforcement learning, which is like machine learning, and also deep learning. It’s more closely related to how a human operates. When you learn how to ride a bike, you use reinforcement learning because when you fall down you learn to avoid what caused that and in this way you slowly build your own neural model for bike riding. Your model gets better and better as you ride the bike, and eventually you may even learn how to do spins in a half pipe. You don’t just learn from data, you learn from simulating the process because it’s more human-like. People are more attracted to this- at least the ones that work with me. People have come to me when they’re more interested in empowering the human and learning from humans- and not just learning from data that humans generate.
AI: Do you have advice for future students looking to work with Computer Science and AI?
PM: Well, you should do it! AI is the future, it really is. Of course, it’s not the only future. Cybersecurity is another really important topic. In my opinion it’s just as important. I’m not saying everyone must do AI, that would be really terrible because we’d forget everything else.
I think anywhere you go, AI is there. Even if you don’t do AI specifically, you should at least know about it and learn about the tools, the techniques, and what the philosophy is behind it. Everyone should know it- it’s just a matter of being open and being educated nowadays. Take a class, maybe do a project, and maybe you’ll do a little bit of AI, even if you don’t do AI directly. I encourage everyone to know it. There’s no way around it.