Research
Machine learning is obsessed with accuracy, but traditional science embraces unexpected results as sources of new insight. My work investigates how errors in carefully engineered ML tasks can teach us about hidden structures in data. I’m generally interested in mixture models, high level data fusion, and stability to distribution shift - usually through the lense of causality.
Topics
Statistical prediction models are often trained on data that is drawn from different probability distributions than their eventual use cases. My (upcomming) work uses insights from causal inference to develop new methods for building machine learning models that are robust to environmental changes.
Interventional distributions acted on by a universal unobserveed confounder can be though of as a mixture model. Cardinality assumptions allow identification of within-component probability distributions, allowing access to inteverntional distirbutions that were previously considered unidentifiable.
We have developed “Expert Graphs” to study consistencies and inconsistencies in partial, but overlapping expertise. As it turns out, this problem is deeply related to issues in voting theory, such as the Cordorcet Paradox.
We process non-coding regions of the genome which contain duplication and mutation signatures. These mutation profiles have been shown to be predictive of various forms of cancer.
Recent Publications
(2023).
Causal Inference Despite Limited Global Confounding via Mixture Models.
In CLEAR 2023.
(2021).
Glioblastoma signature in the DNA of blood-derived cells.
PLoS ONE 16(9).
(2021).
Source Identification for Mixtures of Product Distributions.
In COLT 2021.
(2021).
Synthesizing New Expertise via Collaboration.
In ISIT 2021.