The Machine Learning Center at Georgia Tech (ML@GT) Speaker Series draws leading researchers from across academia and industry each semester to present on current topics in machine learning and artificial intelligence, applications for the technologies, and related insights and experiences. The popular series averages more than 100 attendees for each talk. This spring, ML@GT invited nine experts to campus.
Maya Gupta, a principal scientist at Google AI where she leads the Glassbox Machine Learning Research and Development team came to campus this spring to give her talk, “Constraining Learning Models for Understandability, Regularization, and Fairness”.
Prior to joining Google in 2012, Gupta was a tenured associate professor of electrical engineering at the University of Washington. Gupta has also worked at AT&T Labs, Microsoft, Ricoh Research, Nato’s Undersea Research Center, and Hewlett Packard Research and Development. In 2009 she founded Artifact Puzzles, the second largest U.S. maker of wooden jigsaw puzzles.
Gupta has received numerous awards including the PECASE award and the 2007 Office of Naval Research YIP Award.
Gupta holds a Ph.D. in electrical engineering from Stanford University where she worked with Bob Gray, Rob Tibshirani, and Richard Olsen and was a National Science Foundation Graduate Fellow.
While she was in Atlanta, we had the opportunity to talk with Gupta about what inspired her to have a career in tech, what aspect of future AI excites her the most, and more.
What first drew you to AI and a career in tech?
I loved poker as a kid, because it was all about making good guesses and probability, and that’s really what ML is about too: teaching computers to gamble well.
What are some of the projects that you are currently working on?
My team focuses on how to control machine learning better, so it works like we want it to, and fails in more predictable and reasonable ways. One way we do that is by constraining the shape that nonlinear functions can take.
For example, our latest work trains models that can act on a set of feature vectors, rather than just one feature vector (like if you want to predict something about a product from the set of its customer reviews) and then we show how to learn outputs that generalize the idea of weighted means.
Tell me about how your companies Artifact Puzzles and Ecru Puzzles. How did you get into making puzzles?
I hadn’t done a jigsaw puzzle in years, but I was at the airport heading to Hong Kong and I picked up a puzzle as a gift for a friend’s kids. But then the 14-hour flight was so long and boring, I ended up doing the puzzle myself! It was so fun I looked into doing more jigsaw puzzles, but couldn’t find any images I liked, so decided to start my own jigsaw puzzle company.
Around that time, laser cutters were pretty new, and starting a puzzle company seemed like a great excuse to buy myself a $10,000 laser.
What advice would you give to someone going into this field, especially a woman?
Some advice to anyone, but I think women especially struggle with this, is to figure out what the game is that is being played, and make sure you’re playing the right game, or at least choosing not to.
For example, at school, the game is pretty obvious: do your homework and do well on tests, and you’ll get good grades. But in the real-world, whether it’s getting tenure at a specific university, or thriving at a start-up or a corporate research lab, it can be a lot less clear how to “score points”, and you have to ask the senior people explicitly. Don’t just assume you know what will be valued. Oddly, sometimes it’s not even the work itself that is most important. Sometimes the most important thing is letting people know what you want from them, be it a promotion, a specific project, or even just a raise.
I also encourage people to really listen. There are often people around you that have great ideas, but don’t know enough about the problem to even express their idea in the right way. It’s the non-experts who often ask the best questions that cut through the assumptions senior people have gotten too used to.
What is a lesson you have learned from being a lifelong researcher and student?
I’m a big believer in simulations, and the simpler the simulation you can do, the better. Once back at Bell Labs I spent weeks trying to prove something was true, when a quick simulation would have shown me it wasn’t. Simulations are also a great way to test your algorithm ideas, and to gain intuition about what’s happening.
Why do you think it’s important for people to come and give talks like the one you gave today?
We’re all smarter when we learn from each other. I love getting out to universities like Georgia Tech and learning from my colleagues there, as well as letting you guys in on the cool stuff we’re doing at Google AI.
What is something that excites you about the future of AI?
The most dangerous thing most of us do is drive, and I am very much looking forward to a world where AI keeps cars from crashing into each other.
Who is someone that inspires you and why?
Ingrid Daubechies, Voltaire, Agatha Christie, Steven Pinker, Corinna Cortes, Eddie Izzard… I’ll let you figure you figure out why for yourself!
Tell me about something that brings you joy.
I love brainstorming. It’s the perfect mix of creativity, socializing and principled reasoning, and sometimes it’s even useful!
I also love seeing the ocean and all its immensity, because it reminds me how small and insignificant my daily stresses are in the cosmic scheme of things.