Team: Emma Drobina, Shusen Liu, Peer-Timo Bremer
Machine learning models for scientific applications often learn complex, high-dimensional functions that are difficult to understand on their own. Additionally, many existing techniques for evaluating models fail to handle scientific models effectively, because the model can achieve high global accuracy while retaining errors at key extrema. To fill this gap, a team at Lawrence Livermore National Lab led by Shusen Liu and Peer-Timo Bremer designed a method to reduce complex models to two dimensions so that they could view the function learned by the model as a contour map, which highlights the extrema, saddle points, and slope.
During the summer of 2020, I joined the team as an intern to take the existing code for NDDAV and make it available as a Jupyter Notebook extension. This involved familiarizing myself with the Jupyter Notebook extension creation process (including the use of ipywidgets and traitlets), learning how to launch and communicate between additional processes from Jupyter Notebook, and developing my fluency in Javascript and Python/Javascript communication.
The branch containing my code for NDDAV can be viewed here on GitHub.
I completed the Jupyter extension with time to spare and presented my results at LLNL’s Summer SLAM and at the Lawrence Livermore Foundation’s special edition Summer SLAM, where I presented to LLNL retirees. This was a great opportunity to improve my scientific communication skills and get direct feedback on my presentation. My slides can be viewed below:
