Left: Cathin; Right: her fellow intern

Can you tell us about yourself and your masters topic?
I’m a masters student at Stanford University, where I’m a part of the Computational

Vision and Geometry Lab — I actually just joined this October, and I’m working on projects related to semantic segmentation. I also studied at Stanford as an undergrad, and previously I worked under Sebastian Thrun with Andre Esteva and Brett Kuprel on deep learning for skin cancer detection. So I work on a lot of vision projects, and I’m especially interested in projects that lie at the intersection of machine learning and healthcare. I’m also really interested in human cognition! I loved reading books by Oliver Sacks and other neuroscientists as a kid, but when I first started in computer science, I never considered that there would be much of a direct overlap where I’d get to actually mess around in both fields. Within artificial intelligence research, though, it seems like we still have a lot to learn from actual human brains.

How did you get to work in this area?
There’s this class at Stanford, CS231N, on deep learning for computer vision. On the very first day, I remember that the professor who co-taught the class — Dr. Fei Fei Li — went through this presentation, and one of the slides was about how the initial layers of convolutional neural networks learn basic edge detecting filters that actually closely parallel the basic edge detectors found in cat and human visual cortices, suggesting that there was a deeper and more fundamental connection between these two vision systems. I thought that was insane, and also insanely cool. I joined Sebastian Thrun’s lab a little later, and have been working on AI research since then.

Why did you apply for an internship at Google and how supportive was your masters advisor?
I’d heard really great things about research at Google, and even in my classes and labs, read lots of very impressive work coming from teams in Mountain View, London, and Zurich. I was hoping to get a better sense of what research looks like outside of an academic setting, and the scope of projects and expertise was a huge draw. Also, zillions of GPUs.

My master’s advisor at Stanford is Dan Jurafsky, who is the man. He’s a computer scientist and linguist, has written a book about the language of food, and is basically the best, as far as I’m concerned. He was super supportive.

What project was your internship focused on?
I worked under Andrea Gesmundo from the Applied Machine Intelligence team on Multitask Neural Model Search, a framework to automate deep learning architecture design using reinforcement learning. This work builds off of the Neural Architecture Search research done by Barrett and Quoc from the Google Brain team - that framework was one of the first to successfully apply reinforcement learning to automatically generate convolutional neural networks.

Our project focused on extending that framework so that we could automatically design architectures for multiple different tasks, simultaneously. For example, the same framework could design a model that worked well for sentiment analysis tasks, and another that worked well for language identification, at the same time.

We then showed that it was possible to transfer learn that framework, so that knowledge learned from designing architectures for previous tasks could be reused in totally new, unseen settings. When actual humans design machine learning models, we don’t start completely from scratch every time — we can take advantage of general intuitive design patterns we’ve observed before, as well as remember what models did and didn’t work on similar tasks in the past — and this research tries to take a step closer to doing the same thing in our automated model design.

Did you publish at Google during your internship?
Yes! We submitted our work to ICML, where it’s currently under review (so fingers crossed). The pre-print is also up on Arxiv.

How closely connected was the work you did during your internship to your masters topic?
Although Andrea and I discussed a bunch of project ideas in the months before the internship, this project was actually a chance to try something fairly different from my master’s research at Stanford. For me, at least, that turned out to be one of the best things about this internship — I really loved the chance to explore a very different aspect of AI research, especially one that benefited from the guidance and computational resources available within Google, and I left with a much deeper interest in reinforcement learning that I’ve continued to explore back at Stanford.

Did you write your own code?
Heck, yeah! And then I deployed it cavalierly with enormous care across tons of GPUs. One really awesome thing about interning, though, is the chance to build off of the collaborative engineering effort of other incredibly talented engineers and researchers. I worked pretty closely with code that was being updated almost daily by researchers on the Brain team over in Mountain View, and that kind of cross-continental engineering work feels really neat.

This is your third internship at Google. What were the reasons to come back to Google Zurich?
Third time’s the charm? But actually, I’ve been lucky enough to work at a different office, on very different projects, during all three internships at Google — after my freshman year, I worked with the Glass team in Mountain View, and later I worked in New York on Google Classroom. Each time, I left with a much deeper understanding and appreciation for that particular field, and the care and expertise each of those teams brought to those particular domains. This summer, though, I wanted to come back to work on research in particular. Both of my previous internships had been very software engineering focused, and I was excited to work on AI research that more closely parallels the work I’m excited about at Stanford.

Also, Zurich! I’ve never been to Switzerland before, and this summer one of my fellow interns and I took a train out to hike past the Matterhorn. She wisely remembered to bring along a Toblerone bar for comparison. The real thing is much more breathtaking (but a lot less chocolatey.)