Lester Mackey

About Lester’s Work

Lester Mackey is a computer scientist and statistician advancing solutions to data science problems with practical applications. Mackey’s research in machine learning and statistics focuses on techniques to improve efficiency and predictive performance in computational statistical analysis of very large data sets. He applies his theoretical insights to develop scalable learning algorithms with direct benefit for society.

In early work, Mackey designed a method for more accurately predicting disease progression rates in patients with ALS, or Lou Gehrig’s disease. Working with a newly released database of past clinical data, Mackey and his team used a Bayesian method with feature selection and regression trees, which provided more accurate predictions of prognoses than clinicians had made. Mackey has also addressed subseasonal climate forecasting (two to six weeks out), a more challenging form of weather forecasting due to oceanic and atmospheric variables that can have long-term ripple effects. Mackey’s team created a novel dataset from several existing sources and developed a multi-task learning approach. They identified relevant historical weather patterns using a nearest neighbor search algorithm (which classifies data based on similarity), eliminated historically irrelevant data using multi-task feature selection, and trained two nonlinear regression models to predict future patterns of temperature and precipitation. Their models performed better than the U.S. Bureau of Reclamation’s forecasting results on both temperature and precipitation. More recently, Mackey and colleagues have employed adaptive bias correction on dynamical weather models, further improving subseasonal forecasts that are critical to decision making on matters such as resource allocation, flood mitigation, and wildfire management.

Mackey’s contributions to statistical theory include a new procedure for compressing probability distributions. Data compression identifies a subset, or sample, of data that represents the critical information of the entire dataset. Mackey and a collaborator created a kernel thinning algorithm that enables more efficient processing and analysis and that can be scaled to very large sample sizes, such as those in medicine. Mackey also extended Stein’s method (first published in the 1970s) for use in multivariate problems and for a sample data quality measure that bounds the discrepancy between sample and target expectations. Inspired by real world problems, Mackey pioneers innovative statistical and computational techniques for the common good.

Biography

Lester Mackey received a BSE (2007) from Princeton University and an MA (2011) and a PhD (2012) from the University of California at Berkeley. He was an assistant professor in statistics and created the Statistics for Social Good working group at Stanford University (2013–2016) before moving to Microsoft Research in 2016, where he is currently a principal researcher. His papers have appeared in Nature Communications, the International Conference on Machine Learning (ICML), Advances in Neural Information Processing Systems (NeurIPS), the Annals of Statistics, the Annals of Probability, and the Annals of Applied Probability.

Published on October 4, 2023