On April 24, the Center for Statistics and Machine Learning (CSML) and the Princeton Institute for Computational Science and Engineering (PICSciE) teamed up to host their first ever joint graduate certificate colloquium. Throughout the afternoon, students enrolled in the graduate certificate programs in Statistics and Machine Learning (SML) and Computational Science and Engineering (CSE) presented their research. The interdisciplinary group of students discussed their findings as they relate to statistics and machine learning and computational science.
“Computing has an impact on basically every discipline,” said Michael E. Mueller, associate chair and professor of mechanical and aerospace engineering and director of the graduate certificate in computational science and engineering.
Professors Michael E. Mueller and Tom Griffiths giving welcome remarks during the joint colloquium.
Before the presentations started, Mueller noted that the program purposefully didn’t identify which presenters were from the SML certificate program versus the CSE certificate program. “That was on purpose,” said Mueller. “So that the audience can see how similar the two really are in utilizing computing for a wide range of activities.”
“The overlap between SML and computational methods is increasing,” echoed Tom Griffiths, professor of psychology and computer science and director of CSML. “It’s a chance for people to have their horizons broadened.”
Impact across disciplines
Eight graduate students from six different departments were granted 20 minutes each to discuss their research at the colloquium. From plasma physics to neuroscience, each of the students discussed the ways in which machine learning and computational science helped forward discovery in their work.
PICSciE graduate certificate student Joseph Lockwood and a figure from his research “Modeling Tropical Cyclone and Coastal Risk in Changing Climate: Machine Learning, Correlations, and Resilience.” Advisor: Michael Oppenheimer.
From the Department of Geosciences, Joseph Lockwood presented on short-duration, high impact hurricanes. Lockwood used machine learning to develop models of the rain shape and rain rate intensity of tropical cyclones. He trained the model on a set of 26 historical tropical cyclone events and found that in the end it could accurately capture the predicted rainfall of an event. “The model worked really well,” said Lockwood. (In a convenient, real life illustration of rain fall, a localized rain shower occurred during Lockwood’s presentation.)
CSML graduate certificate student Marie-Lou Laprise and a slide from her presentation “Making Language Models Smarter: Leveraging Legal Logic for Enhanced Reasoning Capabilities.” Advisor: Brandon Stewart.
Marie-Lou Laprise from the Department of Sociology is evaluating the reasoning capabilities of large language models. With new LLMs constantly coming out, “It’s really hard to assess what they’re capable of,” said Laprise. Enter the benchmark test LEGAL-MIX, with which Laprise is evaluating the abstraction capabilities of LLMs by seeing how well they can follow legal fact patterns. Laprise said with a lot of benchmark tests, LLMs could get answers right – but for the wrong reasons. LEGAL-MIX is a way of testing the shortcomings of chatbots. Laprise said she’s interested in future work studying whether improving the legal reasoning of LLMs improves their reasoning in other areas, such as mathematics.
PICSciE graduate certificate student Charles Maher and a figure from his research “Hyperuniformity of Maximally Random Jammed Packings of Hyperspheres Across Spatial Dimensions.” Advisor: Salvatore Torquato.
In the Department of Chemistry, Charles Maher is using computational methods to study a state of randomly configured, mechanically rigid, non overlapping objects. This state of maximally random jammed packing is hyperuniform – meaning there’s a suppression of large wavelength density fluctuations as compared to typical liquids. Maher’s work explored the importance of the link between strictly jamming and hyperuniformity in the packing.
PICSciE graduate certificate student Tal Rubin and a figure from his research “Solving Multi-Fluid Equations in Curvilinear Coordinates, Combining Conservative Methods with Metric-Induced Sources.” Advisor: Nathaniel Fisch.
CSML graduate certificate student Luther Yap and a figure from his research “Asymptotic Theory for Two-Way Clustering.” Advisors: Michal Kolesar and David Lee.
CSML Graduate Certificate student Nick McGreivy and a figure from his research “Weak Baselines and Reporting Biases Lead to Overoptimism in Machine Learning for Fluid-Related Partial Differential Equations.” Advisor: Ammar Hakim.
CSML graduate certificate student Jamie Chin Chiu and a figure from her research “Examining Variability in Depression, Anxiety, and Quality of Life Between Matched Samples of Treatment Seekers and Non-Seekers.” Advisor: Yael Niv.
PICSciE graduate certificate student Dingyun Liu and a figure from her research “Implicit BGK Collision Operator for Gyrokinetic Simulation with Moments-Conserving Maxwellian.” Advisor: Greg Hammett.
Presenters Tal Rubin, Luther Yap, Nick McGreivy, Jamie Chin Chiu, and Dingyun Liu additionally gave talks about their use of computational methods for understanding questions in plasma physics, economics, and psychology.
“It was wonderful to see the range of topics that students pursuing the Statistics and Machine Learning graduate certificate have been working on,” said Griffiths. “I learned a lot from the talks by the Computational Science and Engineering students.”
“Computing has become an invaluable tool across all disciplines as evidenced by a fantastic set of presentations,” said Mueller. “Common themes, approaches, and algorithms are used, and events such as the joint colloquium allow for exchange and diffusion of ideas across seemingly unrelated disciplines.”