 |
Center for Nonlinear Studies |
The adoption of machine learning (ML) by domains like computational science and cyber-physical systems has brought immense opportunity to accelerate research progress in these disciplines. However, idiosyncratic domain properties like high data generation costs, data corruption or compute restrictions often lead to novel failure modes of otherwise successful ML pipelines. This talk proposes a new design paradigm called 'parsimonious machine learning' as a viable solution to overcome common ML failure modes due to data and compute paucity in scientific and cyber-physical contexts. Specifically, we will discuss novel techniques that develop generalizable ML pipelines, alleviating the adverse effects of data paucity and compute paucity by leveraging a 'hierarchy-of-compute' design.
Bio: Nikhil Muralidhar is an Assistant Professor in the Computer Science Department at Stevens Institute of Technology. At Stevens, Nikhil leads the Scientific AI (ScAI) lab with a focus on developing novel machine learning techniques for application in scientific, cyber-physical and epidemiological contexts. Nikhil has published over 25 articles in top-tier conferences like ICLR, IJCAI, AAAI, IEEE ICDM, SIAM SDM and journals like ACM TIST and Physics of Fluids. He has also contributed a book chapter to the book titled: Science Guided Machine Learning: Emerging Trends in Combining Scientific Knowledge with Data-driven Methods. Prior to joining Stevens, Nikhil completed his Ph.D at Virginia Tech and was nominated for College of Engineering Outstanding Doctoral Student award by the Computer Science Department, as well as nominated for the Outstanding doctoral dissertation award. Nikhil was also awarded the outstanding academic achievement award during his M.S. at George Mason University.