Machine Learning for the Sciences

Written by
Sharon Adarlo
Feb. 21, 2020

In late January, a packed house of 40+ people gathered at Princeton University’s Center for Statistics and Machine Learning (CSML) to hear Sam Greydanus, an AI resident from Google Brain, give a presentation on his latest findings concerning neural networks. Greydanus’ talk delved specifically into teaching neural networks physical laws by using Lagrangians, a mathematical description of a physical system. 

Since the fall, CSML has been playing host to these meetings, a series of lunch and learn Friday sessions dubbed Machine Learning for the Sciences (also called the Machine Learning Lunch), informal gatherings that have featured speakers within Princeton and outside practitioners like Greydanus to discuss how they are using machine learning in their research work.

The rapt crowd at Greydanus’ talk hailed from a multitude of departments on campus – including chemistry, physics, astrophysics, electrical engineering and computer science. The mix of faculty, postdocs, and graduate and undergraduate students was typical of these gatherings.

“The idea is to create a surface for people to interact,” said Peter Melchior, an assistant professor jointly appointed to the Department of Astrophysical Sciences and CSML, who started the meetings. “On campus, we have strong expertise in certain types of machine learning in engineering that have become relevant in physics and vice versa. The meetings create an exchange of ideas between these different domains.”

“Peter Melchior’s Machine Learning for the Sciences exemplifies an important part of our mission: fostering interdisciplinary learning and connections,” said Peter J. Ramadge, CSML director. “The sessions have been interesting, informative and very well attended.”

Melchior started the meetings a year ago at the astrophysical sciences department where he worked previously as a professional specialist. They started off as lectures on data science topics catering to astrophysics students and have changed into a mix of tutorials and invited speakers.

When he was appointed to CSML last fall semester, Melchior took the opportunity to broaden the scope of the gatherings so that the subject matter covered could be widely applicable to other science and engineering disciplines, he said.

“You see how methodologies can be deployed elsewhere, that a capable method can be utilized in a different domain,” continued Melchior. “There is familiarity and common ground in that sense.” 

Another important aspect of the meetings is that they provide a window for researchers to see how other scholars talk about their work.

“Sometimes, we have to do this translation work, so we can see how people phrase the work they do,” he said.

Recent meetings have featured Michael Churchill, a research scientist from the Princeton Plasma Physics Laboratory, who gave a talk on using neural networks to model fusion plasma reactions within tokamak nuclear reactors, a device that uses magnetic fields to confine hot plasma into a torus, the shape of a doughnut.

In another meeting, Melchior presented a technique, a Bayesian machine learning architecture, and used examples from his own research to show this method in action. Melchior’s research focuses on developing sophisticated machine learning algorithms to extract data from large astronomical surveys. An article about his work can be read here. 

For more information on these meetings, contact Peter Melchior at [email protected] and also go here. More on Melchior’s work can be found on his website.