Computation’s summer hackathon, part of a thrice-yearly series at Lawrence Livermore National Laboratory (LLNL), kicked off on July 13 in the High Performance Computing Innovation Center (HPCIC). Participants ranged from summer students and new hires experiencing a longstanding LLNL tradition for the first time, to repeat hackers who attend the event each season. With the freedom to pursue any task relevant to their work for Computation, employees experimented with new tools and languages, developed machine learning applications and explored efficiency projects.
Figure 1. In continuation of a spring hackathon project, Huy Le (standing) demonstrates his work with HoloLens augmented reality goggles. (Photo by Isabel Futral.)
The hackathon began with an introduction from Computation’s new Associate Director (AD) Bruce Hendrickson. He affirmed the value hackathon projects provide for LLNL, noting their entrepreneurial spirit. He also expressed appreciation for the fact that hackathons “bring together people who don’t normally work together,” allowing for unique collaboration. This was the first hackathon Hendrickson had the opportunity to attend after accepting the AD position in February, and his presence did not go unnoticed by participants. “By having senior management present for the event, which in the summer includes many new hires and summer students, there is an air that ‘this work you’re doing is important and noticed,’” says Ian Lee, who co-organized the event with Greg Becker, both from Livermore Computing (LC).
Lee, a frequent hackathon attendee and repeat member of the organizing committee, cited the diversity and interactivity of the hackathon as one of the primary reasons he continues to participate. “Hackathons are a great opportunity to connect with and share knowledge between the various divisions and groups across Computation,” he explains. In addition to organizing responsibilities, Lee was a member of an eight-person team that embodied the diversity of thought that the hackathons inspire. The group consisted of employees from different areas in Computation including LC, Applications, Simulations, and Quality (ASQ), and Livermore Information Technology (LivIT), and with roles ranging from developers to operations personnel to new hires. “This combination of talent allowed us to share some interesting technologies and techniques. None of us was familiar with all of the tools, and everyone learned about at least one new concept or skill,” says Lee.
Figure 2. Kyle Dickerson (left) and summer student Scott Lim discuss their hackathon project on gamifying data entry. (Photo by Isabel Futral.)
ASQ scientists Joe Eklund and Daniel Gardner partnered to create a tutorial for TensorFlow, an open source deep learning library created by Google, and Keras, a neural network library written in Python that is compatible with TensorFlow. Used together, the tools can quickly create machine learning models. With a basic understanding of TensorFlow’s capabilities, Keras allows for a higher-level usage to solve problems faster. “Think of it like when you are taught how to do long division as a kid, but now that you know how to do it you still just use a calculator,” explains Eklund. While Gardner was familiar with TensorFlow through his coursework, Eklund still thought of some of the math behind machine learning as “seemingly black magic.” By teaming up, Gardner was able to test his knowledge of the tools, while Eklund was exposed to new libraries.
One team comprised entirely of summer interns worked together on an efficiency project. Undergraduate students Cynthia Lai, Vivek Ramchandran, Karan Shah and Rishi Veerapaneni worked on a web application to streamline the process of calling an LLNL Taxi—a service provided to employees for rides around the main campus. The inspiration for their project came from lunch time, where Lai jokes, “Interns are notorious for calling taxis.” As more students would join in the wait for a taxi, the ones that arrived would often be too small, and drivers had noted how many more requests they received with interns on campus. The hackathon team utilized a Google Form for users to input location, destination and size of their party, and built a clobbering feature using Python script to find the most efficient route. They also included front-end user features, such as advance scheduling, although the 24-hour time period limited the complexity of their web application. Three of the four team members had participated in hackathons at their respective colleges, but Shah notes the different environment at LLNL’s event: “This event wasn’t competitive, so we could spend the whole time focused on learning.”
Figure 3. Drivers and dispatchers of LLNL’s Taxi fleet see an uptick of usage in summer months, when the intern population rises. A hackathon team explored rider-request efficiency via a web application. (Photo by Jacqueline McBride.)
Global Security scientists Aram Avila-Herrera and Dayanara Lebron-Aldea wanted to experiment with new tools they do not encounter much in their day-to-day work. Both were first time hackathon participants who focused on Natural Language Processing (NLP). For their project, they used different Python libraries to analyze and compare pairs of questions on Quora—for instance, “how do I make friends” and “how to make friends.” Lebron-Aldea used the Natural Language Toolkit, an open source suite, to differentiate and match questions through lemmatization and stemming. Both techniques attempt to identify common forms of a word by removing affixes or inflectional endings, though lemmatization does so by taking the form into account while stemming simply removes the end of the word with less accuracy. Lebron-Aldea found that lemmatization and stemming decreased accuracy, revealing the importance of training a computer in syntax and grammar for such a task.
Avila-Herrera used the Gensim library and a GloVe model, which is a learning algorithm developed by Stanford for the purpose of converting words to vectors and examining word distribution. His results with the GloVe model illustrated that duplicate questions overall have fewer dissimilar phrases than duplicate questions, a first step in distinguishing individual questions. Despite initially running into problems installing the software on his laptop, the collaborative spirit of the hackathon and 24-hour time limit allowed Avila-Herrera to find a quick solution. “It was a friendly refresher of how important it is to ask for help and to pivot from roadblocks when on a tight deadline,” he adds.
Figure 4. Aram Avila-Herrera (left) and Dayanara Lebron-Aldea discuss the importance of syntax and grammar in the context of NLP after finding that lemmatization and stemming decreased the accuracy of their results. (Photo by Isabel Futral.)
For Computation employees, the hackathons are a vital tradition because of the creativity they inspire, which lasts beyond the 24 hours of the event. Both Hendrickson and LC division leader Becky Springmeyer attended the project presentations at the close of the hackathon, and reaffirmed management’s support for the event. In addition to the cross-directorate collaboration hackathons create, they also provide opportunities for grassroots ideas to be incorporated into Computation’s programmatic work. “These can often be ideas that fit just outside of normal work duties, and the hackathon is a low-risk, high-reward venue for investigating those ideas,” says Lee.
Repeat attendees and newcomers alike can look forward to the fall hackathon, which will be scheduled for October or November.
– Isabel Futral, ISCR