Trustworthy Machine Learning for data-driven discovery

Modern machine learning techniques have ushered in an era of data-driven astronomy.  Flexible new ML tools enable powerful data analysis strategies that were not feasible even a few years ago. This shift in our scientific approach requires us to ask an important question:  Can we trust the black box?  In this talk, I will highlight opportunities and challenges for creating credible ML to interpret galaxy cluster observations.  I will expand on ML interpretability and domain adaptation as keystones for building models that give trustworthy results.  And I will show examples of how machine learning can be used, not just as a tool for getting “better” results at the expense of understanding but also as a partner that can point us toward physical discovery.

Michelle Ntampaka is a Data Scientist and Assistant Astronomer at Space Telescope Science Institute. Her research develops new ways to interpret observations of galaxy clusters and other large scale structures. She uses machine learning and statistics to tease out complicated patterns in the data that are inaccessible through more traditional means. Her physics Ph.D. – Cosmology with Galaxy Cluster Dynamics Using Machine Learning and Forward Modeling – is from Carnegie Mellon University in Pittsburgh, PA.

A former teacher, Michelle puts her experience to good use at every opportunity. She has traveled to Africa to train Rwandan high school science teachers on delivering memorable lecture demonstrations, and was able to enjoy Kigali’s solar eclipse in 2013. More recently, she has developed lesson plans to teach astronomy at the elementary school level.