Daniela Witten
University of Washington, WA·Mathematics & Statistics
Modern data analysis faces two linked problems: checking that a fitted model reflects real signal rather than noise, and drawing sound conclusions when the exact measurements needed are scarce or expensive. The two projects here develop statistical methods that address each, one by splitting data into independent pieces for fitting and validation, the other by using machine learning predictions to stretch small labeled datasets.
- Fields
Mathematics & Statistics
Computer Science
- Topics
statistics
statistical inference
machine learning
- Grants
- 2 known grants, through Aug 2028
What they work on
- Establish semi-parametric efficiency results showing when, and by how much, extra observations with only features X (no outcome Y) improve inference in semi-supervised learning.
- Develop new and improved estimators for prediction-powered inference (PPI), which combines machine learning predictions with a small set of labeled data to draw valid statistical conclusions.
- Propose PPI estimators of true positive rate, false positive rate and area under the curve, standard measures of how well a classifier or diagnostic test performs.
- Generalize data thinning, which splits a dataset into independent pieces for fitting and validation, beyond convolution-closed distributions using sufficiency, and release software tailored to biomedical data.
More about the research
The first project asks how machine learning or artificial intelligence outputs can supplement limited data while keeping statistical conclusions valid. In a typical setting, a researcher has a few complete observations of both a feature X and an outcome Y, plus many observations of X alone. The team will ask when, how much, and how these extra X-only observations improve inference. They will approach this theoretically, by proving new semi-parametric efficiency results (limits on how precise an estimate can be) for semi-supervised learning. They will develop improved estimators for prediction-powered inference (PPI), a framework that combines model predictions with a small set of true labels. Finally, they will build PPI estimators for true positive rate, false positive rate, and area under the curve, standard measures of how well a classifier or test performs.
Why it matters
Researchers often fit complex models to data and then need to know whether the findings are real, or must draw conclusions from few expensive measurements. Methods for validating models and for safely using AI predictions could help biomedical and other scientists reach trustworthy conclusions, including better evaluation of diagnostic tests through rates of true and false positives.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Richard Guo
University of Michigan, MI·Mathematics & Statistics
Develop a hunt-and-test version of the classical score test for infinite-dimensional parameters such as regression functions and log-densities, using machine learning to find signals and debiasing to keep the test calibrated.
Tianxi Cai
Harvard University, MA·Mathematics & Statistics
Develop semi-supervised risk prediction methods that combine noisy electronic health record surrogates, such as billing codes, with a small set of chart-reviewed outcome labels, even when predictors are numerous.
Armeen Taeb
University of Washington, WA·Mathematics & Statistics
Develop a framework that organizes complex model classes (partitions, permutations, directed acyclic graphs, subspaces) as partially ordered sets, and use it to define generalized false positive error.
The school
University of Washington: 81 labs on PI Pages, $1.7B a year on research, 694 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.