We first met with Swetha Revanur last summer, when she was a recent high school graduate heading for Stanford University and interning in HP’s Emerging Compute Lab on a project that used sensor data to create simulations of how people move around in different living spaces. This year, Revanur is back in the same lab but working on a new challenge. We caught up with her to see how her academic interests have developed over the last twelve months and to learn about what she’s been working on this time around.
HP: First of all, how was your freshman year at Stanford?
I had an amazing freshman year! I’ve met some of the most brilliant people, the classes were just the right amount of challenging, and I joined an acapella group on campus. In December, I also traveled out to Sweden to speak at the 2016 Nobel Prize Ceremonies and meet the laureates. I’m excited to start my sophomore year in September!
HP: Are you still planning to major in computer science?
Yes, that hasn’t changed! When I started at Stanford, I was interested in biocomputation, but my interests have since shifted to artificial intelligence.
HP: What prompted the change?
The decision was actually driven largely by my work at HP Labs last summer where I had a lot of exposure to the algorithmic side of computer science. I think that if I can understand these algorithms and optimize them, I can have a much larger impact in whatever sector I choose to work in. At the end of the day, machine learning can always be applied to health, and it has a huge scope.
HP: So what are you working on this year?
I’m with the same team in the Emerging Compute Lab, but instead of looking at sensor analytics, I’ve shifted my focus to the intersection of deep learning and robotics. I’m using techniques in reinforcement learning, which lets us train software agents to find the optimal actions to take in specific environments. I’ve developed a hybrid approach that maintains the same performance as state-of-the-art reinforcement learning algorithms, while improving data and cost efficiency.
HP: How’s it going?
Reinforcement learning is a new area of study for me, and so it’s been a fruitful process of self-teaching. Initially, I was wrangling with pages of linear algebra to understand how existing methods work. Once I got my bearings, I was able to point out gaps and come up with optimizations, and now I’ve implemented the algorithm in TensorFlow.
HP: How will you test the new algorithm?
The new hybrid algorithm will be tested in simulation. I’ll start with simple tests with basic software agents. For example, I recently ran a test where a pendulum was trained to stay upright. Gradually, we’ll work up to full humanoid simulations.
HP: Why is HP interested in this work?
A lot of folks in HP Labs are working in a fundamental robotics research space, on projects like mapping, localization, and navigation. My hybrid approach helps cut time and cost requirements in that space. In general, robotics dovetails really well into the social, business, and home application layers that HP is a major player in.
I was invited to speak at the HP Labs global all-employee meeting with our CTO, Shane Wall. The implications of better reinforcement learning are broad, the interest is there, and I’m excited to see where it takes us.