Faculty Perspectives: Kyle Bradbury
Bass Connections Project Team: A Wider Lens on Energy: Adapting Deep Learning Techniques to Inform Energy Access Decisions
Kyle Bradbury has led Bass Connections project teams since the program’s inception in 2013. A Duke alumnus for his Ph.D. and M.S., he brings experience in machine learning and statistical modeling to energy problems.
In 2019-20, he is leading a project team along with colleagues Leslie Collins and Jordan Malof of the Pratt School of Engineering. Below are excerpts from Bradbury’s remarks at a Spring 2020 Bass Connections orientation for team leaders.
Where Is the Energy Infrastructure?
One in nine people don’t have access to electricity, which affects things like health outcomes, educational outcomes, inequality and economic opportunities. That leads to a lot of good reasons for trying to extend electricity access as much as possible. But in order to determine how to optimally extend the grid to meet the needs of communities without access, we need to know where the existing infrastructure is, such as transmission lines and distribution lines.
We need this information to be able to determine whether or not the optimal strategy is to provide an off-grid solar PV array or to extend the grid out—and that can be a really expensive process, so first we need to know where the infrastructure is to plan accordingly. That information is not generally available in many developing countries, so how do we go about getting it?
Applying Machine Learning to Satellite Imagery
This year’s Bass Connections project explores how we can apply machine learning techniques to satellite imagery data. Can we look at a satellite image, identify the transmission lines and map those out to see where they connect?
There are a number of challenges, one of which is that it can be really expensive to create the “training” data that we need. For automatic identification of transmission lines, you need to be able to train an algorithm by providing examples of what a transmission line looks like. And we need to scale up these techniques to operate at a national level to do meaningful work in the countries of interest.
So, we’re exploring how to use synthetic imagery in order to supplement our existing real satellite imagery to overcome this issue and be able to apply this at scale over a large area.
Grappling with Uncertainty
Students are learning about electricity access and the needs of the communities that we’re talking about. They’re learning how to implement machine learning algorithms to see if this is a realistic and potentially feasible product. This leads to uncertainty: this technique may or may not work, and sometimes our next steps depend on what we learn along the way. Students have been grappling with that throughout the year. It’s a challenge that puts them into an interesting area outside of their typical comfort zone of knowing there will be an answer.
This year we have three leaders, four faculty contributors who will come and talk with students for one or two sessions and be available as resources, one grad student project manager and seven undergraduates. That ratio works out well. Our undergraduates come from computer science, mechanical engineering, statistics, energy, environment and the MIDS program—so we have a pretty wide range of curricular perspectives.
We typically connect our project to a Data+ project in the summer. The Data+ project lays the foundation, typically building a dataset that Bass Connections students will then use in the academic year. Oftentimes what we find is that the Data+ students want to join the Bass Connections project or vice versa, so there’s a bit of continuity between the two, which is really beneficial for carrying that project forward.
For this project, students will be producing a poster and giving a final presentation to the wider Duke community. There is always the possibility of producing a conference paper based on their work.
In the past, we’ve had students produce journal papers as well as a number of conference papers. The research outputs also include datasets that have been used to enable other publications. We’ve had conference papers published pretty much every year based on outcomes that the students initiated through their work on the Bass Connections and Data+ projects. It’s been excellent for research in that respect.
Students won four poster sessions, and one former project manager was invited to the North Carolina State Capitol to present his work.
Advice for New Team Leaders
- Set meeting times beforehand and let potential applicants know.
- Start with team building to get students talking to each other very quickly. We’ve done an escape room, where students are locked in a room for one hour and have to solve puzzles together to figure out how to get out—it’s more fun than it sounds.
- Meet as frequently as you feel comfortable with (we meet twice a week), and ask students to identify a time on their own to meet outside the team meetings and make progress.
- Consider dividing the two semesters into four quarters. In our project, a mini boot camp transitions the students to a place where they are taking more ownership and pushing things forward. The next two quarters are full effort. Save a quarter for report-out and lots of time for students to put deliverables together.
- Try to help pick up students when they fall. There can be times when they know what they’re supposed to do, but other stuff gets in the way. I’ll say, “Okay, that’s fine, let’s use this as a learning experience and move on from here.” It empowers them to take responsibility.
- Ask for frequent deliverables. With this idea of ambiguity that’s present, any type of structure that can be provided while still giving students the opportunity to work within that uncertainty is really important.
- Give students opportunities to connect their work to the larger community through poster sessions, not just the Bass Connections Showcase but also opportunities throughout the year to give them a sense of ownership and pride in their work as they move through it.
See other faculty perspectives and learn how you can get involved in Bass Connections.