Research
My research develops statistical methods that make machine-learning systems more reliable, informative, and useful for decision-making. I am especially interested in settings where standard point predictions are not enough: complex outputs, hidden targets, distribution shifts, and high-stakes scientific systems.
Conformal prediction & inference
I design distribution-free methods that turn model outputs into calibrated uncertainty sets. Current work studies generative prediction, efficient coverage allocation, and inference for latent distributional parameters that are never directly observed.
Reliable & secure machine learning
I study how learning systems can preserve utility while remaining robust to misuse and changing operating conditions, including model-extraction defense and statistically controlled monitoring.
Scientific AI & decision-making
I combine physical structure, learned representations, and statistical testing to support decisions in scientific and engineering systems where simulation is costly and uncertainty matters.
Interpretable machine learning
My earlier work uses interactive visualization and representation analysis to diagnose models and data, with applications to geospatial deep learning and network analysis.
Current questions
- How can conformal guarantees be transferred from observable outcomes to scientifically meaningful but unobserved quantities?
- How should uncertainty-set geometry adapt to multimodal, high-dimensional, or structured outputs?
- How can physical knowledge and statistical calibration support fast, trustworthy monitoring under distribution shift?
- How should uncertainty be communicated so that people and AI systems make better joint decisions?
Ongoing projects
- Deciding When to Decide: Decision-Focused Testing under Distributional Shift. A framework for testing whether a decision optimized in one domain remains approximately optimal after a context or feature distribution shift.
- Derandomized Multiple Change Point Detection with FDR Control using Distance Profiles. A nonparametric approach to detecting multiple changes in sequences of metric-space objects while controlling false discoveries.
- Asymptotic Behavior of the Maximum Degree Distribution under Graphon Models. Bootstrap inference for extremal network statistics from a single observed graph.
Interested in one of these questions?
Start a conversation