The Herbert and Florence Irving Institute for Cancer Dynamics will continue its seminar series on the topic of mathematical sciences underpinning cancer research. The monthly seminars take place on the second Wednesday of the month, 2:00-3:00 PM EST. The presentations are open to the Columbia community (in person and online) and to researchers outside Columbia (via Zoom).
On Wednesday, November 13th (2:00 PM ET), IICD welcomes Genevera Allen, Professor of Statistics (IICD & Zuckerman Institute), Columbia University. Seminar hosted by Simon Tavaré. The seminar will take place in person in Schermerhorn Hall 603 (Morningside Heights campus). If you wish to attend the seminar remotely, please register using the following link: https://columbiauniversity.zoom.us/meeting/register/tJMtcOCgrz4oHdG8qHyeXjEQTmLjcD1RkhXE
Title: Developing Reliable Machine Learning Interpretations Through Validation and Statistical Inference
Abstract: Machine learning (ML) techniques are widely used throughout science and biomedicine to make discoveries from troves of data. But, are these data-driven discoveries reliable? Many such discoveries are made using statistical models or interpretable machine learning techniques, which provide human understandable interpretations of blackbox systems. Unlike for prediction tasks, however, it is difficult to directly test the veracity of ML interpretations. In this talk, we first present findings from a systematic, large-scale empirical reliability study on popular ML interpretations for both supervised and unsupervised tasks on tabular data. Our results reveal that ML interpretations are generally unreliable, and dramatically less reliable than ML predictions. Given these negative findings, we secondly ask: How can we improve the reliability of ML interpretations? We discuss the challenging task of validation for ML interpretations, presenting practical strategies like stability and generalizability as well as approaches to quantify the uncertainty of interpretations through statistical inference. We conclude with several real examples of how to apply such approaches to make reliable discoveries from large-scale biomedical data.
Bio: Genevera Allen is a Professor of Statistics. She is also an Investigator at the Zuckerman Institute for Mind, Brain, and Behavior and a core faculty member at the the Irving Institute for Cancer Dynamics. Prior to this, Dr. Allen spent fourteen years at Rice University in the Departments of Electrical and Computer Engineering, Statistics, and Computer Science; she was also the Founder and served as the Faculty Director of Rice's data science education center, informally known as the Rice D2K Lab. Dr. Allen’s research develops new statistical machine learning tools to help people make reliable discoveries from large and complex data. She is known for her methods and theory work in the areas of unsupervised learning, interpretable machine learning, data integration, graphical models, and high-dimensional statistics. Her work is motivated by solving real scientific problems, especially in the areas of neuroscience and computational biology. Dr. Allen is the recipient of several honors including a National Science Foundation Career Award, Rice University’s Duncan Achievement Award for Outstanding Faculty, and in 2014, she was named to the “Forbes ’30 under 30′: Science and Healthcare” list. She is also an elected fellow of the American Statistical Association, Institute of Mathematical Statistics, and International Statistics Institute.
