\nProfessor of Computing and Information Sciences
\nRochester Institute of Technology (RIT)
\n<\/span><\/strong><\/h6>\n
![\"\"](\"https:\/\/cdn.vanderbilt.edu\/vu-URL\/wp-content\/uploads\/sites\/193\/2023\/04\/19203038\/7639_Wang_Linwei_01_HR-203x300.jpg\")
<\/p>\n
Date:<\/strong> Thursday, April 13, 2023 Title: Abstract: Bio: Dr. Wang obtained her BS degree in Optic-Electrical Engineering from Zhejiang University (China) in 2005, her M.Phil degree in Electronic and Computer Engineering from Hong Kong University of Science and Technology in 2007, and her PhD in Computing and Information Sciences from RIT prior to joining the faculty of RIT in 2009.<\/p>\n <\/p>\n
\nTime:\u00a0<\/strong>11:45 a.m. Lunch, 12:00 p.m. start
\nLocation:<\/strong>\u00a0Stevenson 5326<\/p>\n
\n<\/strong>The Risk of One-Size-Fit-All Deep Learning in Personalized Medicine<\/p>\n
\n<\/strong>Patient-specific predictive models, personalized to observations from individual subjects, hold important promise for modern medicine. The successes of state-of-the-art deep learning, however, are built on learning a \u201cone-size-fit-all\u201d model optimized from the average statistics of a large training data. In this talk, we examine the risk of this status quo<\/em> from two different aspects: 1) its impact on the accuracy of forecasting across heterogeneous time series, and 2) its potential to create bias for subpopulations underrepresented in the training data. In part-I, we will examine the limitations associated with learning a single one-size-fit-all<\/em> function to describe diverse heterogeneous dynamics, and discuss our recent efforts in a meta-learning framework that learns to adapt a dynamic function to few-shot time-series observations at test time. We will discuss the benefits of this approach in both general time-series forecasting benchmarks, as well as in the application of learning personalized neural surrogates for cardiac simulations. In part-II, we will examine how learning based on an average loss may create spurious correlations that are biased against underrepresented individuals, and discuss how current state-of-the-art solutions \u2013 while successful in benchmarks with \u201cclean\u201d spurious correlations \u2013 may fail in more realistic medical image tasks where the fundamental assumptions underlying many of these solutions need to be re-considered.<\/p>\n
\n<\/strong>Dr. Linwei Wang is a Professor of Computing and Information Sciences at the Rochester Institute of Technology (RIT) in Rochester, NY, where she serves as the Director of the Personalized Healthcare Technology (PHT180) Research Center that consists of over 100 faculty affiliates across nine colleges at RIT. Dr. Wang also directs the Computational Biomedical Lab at RIT, with core research interests centered around statistical inference, Bayesian deep learning, and inverse problems with applications to signal and image analysis in a variety of domains including health, astrophysics, and material design. Dr. Wang is a recipient of the NSF CAREER Award in 2014 and the United States Presidential Early Career Award for Scientists and Engineers (PECASE) in 2019.\u00a0 Dr. Wang currently serves as the Executive Secretary on the Board of the Medical Image Computing and Computer-Assisted Intervention (MICCAI) Society.<\/p>\n