Friday, April 18th
12:00pm - 1:00pm, Zoom/ARB 627
Meeting ID: 957 7525 9475
Passcode: 834911
Minghe Wang
MS in Biostatistics Student
Columbia University
Understanding and Addressing Covariate Shift in Transfer Learning
Abstract:
Covariate shift refers to systematic differences in the distributions of input covariates between training (source) and deployment (target) settings. This presents a significant challenge when combining data from multiple sources or deploying pre-trained models to real-world healthcare applications. In the first lecture of this tutorial, we will introduce key concepts related to transfer learning and domain adaptation, followed by simple solutions for correcting covariate shifts. We will explore techniques for estimating density ratios, including kernel density estimation and histogram-based methods. Real-world examples using the MIMIC-III dataset will demonstrate how covariate shift impacts model performance. The subsequent lectures will build on this introduction with hands-on labs that cover more advanced reweighting approaches, such as kernel mean matching, discriminative learning, and techniques that go beyond simple reweighting. This series will consist of approximately four lectures.
Speaker Bio:
Minghe Wang is a first-year Master's student in Biostatistics. He received his Bachelor's degree in Applied Mathematics, with a minor in Computer Science, from New York University. He will be leading the lectures and labs for this series of tutorials, under the mentorship of Drs. Tian Gu and Ying Wei.
About TRAIL4Health & the Brown Bag Learning Series
TRAIL4Health is a Translational AI Laboratory committed to advancing public health through innovative applications of artificial intelligence and data science. At TRAIL4Health, we host a variety of events to foster learning, collaboration, and the exchange of ideas at the intersection of AI, public health, and medicine.
The TRAIL Brown Bag Learning Series is a weekly informal gathering held on Fridays at noon in Room 627 of the Allen Rosenfield Building. These sessions are an opportunity for faculty, students, and researchers to come together and learn something new—whether it be a dataset, software, new article, or simply to exchange ideas. This informal setting provides a platform for sharing knowledge, stimulating discussions, and fostering collaboration in a relaxed environment.