List of Projects with Descriptions
† Advisor / Supervisor
E. A. Maceda, E. C. Hector†, A. Lenzi†, and B. J. Reich†. A variational neural Bayes framework for inference on intractable posterior distributions. Accepted (Environmetrics).
Manuscript.
Github Repo.
In this paper, my team and I proposed VaNBayes, a neural posterior estimator that estimates hyperparameters
of the variational posterior instead of modeling the full posterior direction.
We believe that doing this could have some computational benefits. This paper
ended up being a very empirical/application-based paper, with many examples of
this approach working well. This was also my first paper ever, so it has a special
place in my heart. 😊
E. A. Maceda, J. Miller, S. Reyes-Roza, P. Hernandez, M. Miller†, R. Sazdanovic†, N. A. Hill, N. Josephs†, and M. S. Olufsen†. Branching angles in pulmonary arterial networks of control and pulmonary hypertensive mice. In preparation.
Manuscript (coming soon!)
Github (coming soon!)
In this project I use statistical shape analysis to identify the differences between hypertensive
and control pulmonary arterial networks in mice. In her thesis, Dr. Megan Miller investigated the
differences between hypertensive and non-hypertensive mice in pulmonary arterial networks. In this
follow-up paper, we investigate the effect that the shape of these pulmonary arterial networks have
on the blood flow. As the statistician on the team, I was responsible for obtaining network statistics,
statistical shape analysis, and conducting statistical inference comparing the two mice groups.
E. A. Maceda, A.-M. Staicu†, and B. J. Reich†. Physics-informed functional data analysis for real-time NOx emission estimation. In preparation.
Manuscript (coming soon!)
Github (coming soon!)
4D-Var is a powerful approach used in Data Assimilation to combine observed satellite data
and data from a differential equation model, but it is computationally expensive.
In this paper, we propose using functional PCA to build an amortized predictor of NOx emissions
that is trained on data from a differential equation model, but only requires satellite observed
NOx concentrations at prediction time.