Q & A with David Fouhey
David Fouhey is an Assistant Professor at the University of Michigan in the Computer Science and Engineering department.
RA: David, it is a pleasure to speak to you again. The last time you were featured in our magazine was in one of Angjoo Kanazawa’s photographs of the ‘Berkeley Crowd’. A lot has changed since then. What is your life like now?
DF: It’s a lot busier. I have many wonderful students now. For two of them, it’s their first CVPR, so I’m looking forward to hanging out at the posters with them virtually and answering all the questions. I’m also looking forward to sitting up and drinking coffee with one of my students. It’s important that we ensure everyone can come to the posters and see stuff and we can talk to everybody. Life has certainly changed a lot since the last time I featured, and it’s changed even more since the first time, which was four years ago!
RA: How many people are you in the lab now?
DF: There are currently eight graduate students and then I have a large number of undergraduates for the summer, which is exciting. Some of them are working remotely. There’s a lot of stuff going on, but it’s really wonderful to work with students.
RA: Did you always want to continue in academia and keep teaching?
DF: Yes, I have really enjoyed teaching, both in the classroom and getting students excited about computer vision and machine learning. I think it’s important not to hoard knowledge in your head. You have to get it out there. It’s really important for people to learn as much as possible and to teach people and welcome them into the field. Machine learning is very exciting now but there are lots of ways that it can go wrong. As people who have been around a while and seen that, I think it’s important for us to teach the next generation. We don’t want to keep on making the same mistakes.
RA: In what way is being an assistant professor different from what you expected?
DF: I have to do many more things than I realized! Lots of very different things. The topic can totally change from one meeting to the next. From talking about the next iteration of a course, to speaking about someone’s results. I switch around lot, which is exciting, because I see lots of new fun stuff.
RA: Do you find that you learn things from your students?
DF: What’s great is that students often have new and fresh ideas. What’s wonderful about computer vision is that we really, as a field, don’t know what’s going on most of the time. It’s very easy for someone to get started and to think of something totally new that you’ve never thought of before in that way. That’s why it’s wonderful to work with a collection of students from all sorts of different backgrounds. It keeps you on your toes and you get to learn all these new perspectives on things. It’s great. And they also help you keep up with reading arXiv!
RA: Do you ever feel overwhelmed by it all?
DF: Do you have moments where you think, “Get me out of here!” and want to be a software engineer in a start-up instead? I mean, definitely, in academia like in grad school you often do have these moments where nothing works and where your paper gets rejected, then your paper gets rejected again, and it’s really hard at times. Especially when you first start out. You go into this field where the default response is often no. I think it’s very important as a field that, especially as we’re growing, we treat people with respect and actively try to be inclusive of new people. It’s hard enough for my students when their papers get rejected, but at least they have someone who can say, “I’ll fix this,” but when people are just getting started and don’t have mentors floating around in their life, it can be tough. This is a problem that exists when a field grows really quickly, but in the long run the growth is really exciting.
RA: Now that you see the world through the eyes of a teacher, are there things that you see that really aren’t working, and you think we should fix to make the community work better? Funnily enough, last year, I interviewed Andrew Fitzgibbon from Microsoft and I asked him a similar question, and he told me: “Someday we’re going to have to figure out how to do these conferences without everybody travelling to the same place.” Last year, it sounded impossible, but what a difference a year makes!
DF: I really appreciate you asking this question. One of the things that has really changed since I started in computer vision is back then you looked for the example where your system worked and you were really excited, but it was a total fantasy. Like, “Maybe one day my system will work.” Now, we have systems that do stuff. One thing I try to teach, and I want to teach better, is that if you deploy these systems in the real world, if you’re not careful, they can have real consequences. There are all these stories that float in the community from ages ago about data bias. Like an entertaining story about a tank classifier that gets 100 per cent accuracy because it determines whether it’s taken at night or during the day with pictures of Soviet tanks at night and US tanks in the day. But now there are real serious issues where people deploy things. There’s this great paper from Joy Buolamwini and Timnit Gebru on Gender Shades and it has had real downstream impacts. It’s something that as a community we have to start thinking about because we know how a lot of these systems work and we need to make sure they’re not misused. We need to make sure that there aren’t bad outcomes and consequences. There’s an excitement about stuff working, but then this stuff can have really serious impacts and it’s important that as a community we talk about algorithmic bias and address it.
RA: Do you think the community will hear your call? How do you see things changing in this area?
DF: There are many other systemic issues and there’s a lot of reading that everyone can do. A lot of the issues that you spot in these articles are things that you’ll talk about, but for more simple things where you say, “If I trained a classifier to detect giraffes, maybe it only will pick up on some sort of other correlation.” I think it’s something where we talk about these things as academic examples, and it’s kind of interesting when it happens on MS COCO, but when it happens in the real world, we abstract away the concept that data and algorithms can have bias and forget about it. I think these are really hard problems and we have to find solutions. I don’t have solutions, but I think we have to talk about it and be aware of it and listen to people who have been talking about it for quite some time.
RA: Thinking back to the Berkeley Crowd, what do you miss the most from that time and those people?
DF: I miss ditching work and going off for a hike with my lab mates and taking long extended meals where you discuss anything and everything. Those are times that you should treasure in graduate school because you don’t get as many of them after.
RA: I think every one of our readers can relate to that.
DF: One thing that I love about this community is that you see the same people and you’ve known them over many years. I met Angjoo at ECCV 2012. I was not part of the Berkeley Crowd for a while, but I knew them, I would see them at conferences, we’d hang out, we’d talk, we’d catch up. Now, they’re friends for life, and I’m sure in 20 or 30 years from now we’re still going to be in contact. You make these amazing friends over this really long period of time. It’s great. When you start going to CVPR, you don’t expect it. Then you go again and again and again.
RA: That is a really nice message for people attending their first CVPR. Everyone can build their own Berkeley Crowd.
DF: Yes, they’re friends you don’t realize you have yet.
RA: Do you have a funny story from those days that you could share with our readers?
DF: I remember when Alyosha Efros would take us on a hike, he’d say, “It’ll be an hour,” and it’d always be like four hours! We would do things like there was a miniature train that he would take us on, and somehow, we’d always end up eating gelato. He had this uncanny ability to find gelato! These hikes would be outrageously long, and his estimates would be wildly inaccurate, but they were so much fun. I’d come home totally sunburnt but very happy! My message to people is make sure you take the time to do stuff like this because it’s really important.
RA: By having a career in academia, is that your way of not abandoning that world completely?
DF: Yes, I get to talk to people about all sorts of research problems all the time. I can work on all sorts of things. I‘m in heaven! I’m trying to do lots of different projects at the same time and it’s so much fun getting to have that experience with my students. An advisor-advisee relationship is not the same as you and your office mate, but there are similarities. You sit in the office and say, “What problems should we be solving?” Or, “Did you see this new thing on YouTube? How can we use that for computer vision?” It’s wonderful.
RA: Computer vision technology is evolving so fast. Where do you see things going next?
DF: People are really interested in 3D now, which is great. I got interested in 3D when it really didn’t work. Some of my old results are just horribly embarrassingly bad! It’s exciting. Because of deep nets now there’s stuff that you just couldn’t imagine. Justin Johnson is also at Michigan and he’s interested in 3D, so we have two students who we co-advise and it’s a lot of fun.