Research

Broadly, I am interested in questions of trust in a machine learning analysis. I work on tools to detect lack-of-robustness in an ML analysis, and on methods for “robustifying” one: through carefully designed models, algorithms that produce well-calibrated uncertainties, and are robust to poor optima. I am also interested in large-scale foundation models and their ability to encode useful aspects of biology and human physiology.

Foundation models for biology

Models trained on large scale gene expression data promise insights into gene regulation under different conditions and contexts. I work on training these models at scale, on making the architectures cheaper to train, and on the problems that come with predicting responses to perturbations.

See the 4 papers in this thread

Influence functions and robustness

Will the inferences drawn from a particular analysis, or the predictions made by a model, change substantially under perturbations to the training data, minor variations of modeling assumptions, or a different inference algorithm? In this line of research we develop tools for answering those questions.

A naive approach to understanding the effect of data perturbations refits the model to many perturbations of the data. That is infeasible for large datasets and for structured latent variable models, which involve expensive marginalization over latent variables. We develop efficient but accurate approximations that involve a single fit to the dataset, and allow one to perturb data by dropping time-steps from within a time series or sites from a spatial extent. As a bonus, the same machinery approximates cross-validation in hidden Markov models and Markov random fields.

We can also consider the effect of modeling assumptions on the inferences drawn from an analysis. We look at this in the context of Gaussian processes and develop a methodology for measuring sensitivity to the choice of kernel. Somewhat surprisingly, we find that in many cases minor perturbations to the kernel function result in substantially different predictions, calling into question the robustness of the underlying analysis.

See the 4 papers in this thread

Uncertainty quantification and calibration

Quantifying the uncertainty of a prediction made by a modern neural network remains challenging. Yet well-calibrated predictive uncertainties are essential for deciding when to abstain from a prediction in safety-critical applications, for producing diverse outputs from generative models, and for effectively traversing the exploration-exploitation tradeoff.

I am also interested in uncertainty quantification more broadly, and am a core contributor to UQ360, an open source toolbox that provides a number of approaches to quantifying uncertainty, measuring its quality, and communicating it.

See the 7 papers in this thread

Bayesian neural networks and horseshoe priors

Bayesian neural networks hold the promise of retaining the predictive performance of their point-estimated counterparts while providing well-calibrated uncertainties and principled approaches to model selection. My work has explored the effects of commonly used priors and approximate inference algorithms on the quality of posterior uncertainties, and developed algorithms for inference and efficient decision making. My work has explored heavy-tailed horseshoe priors for Bayesian neural networks that shrink almost every unit to zero, but let a few escape shrinkage altogether, allowing model selection to fall out of inference. The resulting models produce better calibrated uncertainties than traditionally used Gaussian priors.

See the 6 papers in this thread

Health and disease progression

Data-driven hypothesis generation can be an effective tool for scientists studying phenomena that are as yet poorly understood, such as researchers trying to understand how neurodegenerative diseases progress. The unique challenges faced in these scenarios have guided my research. Latent variable models are useful for representation learning from clinical registries with noisy and missing data, and more broadly for analyzing case-control studies.

Those representations are useful for characterizing the progression of disease from longitudinal follow-up of patients. Variants of hidden Markov models are effective for describing progression as a sequence of jumps between interpretable disease states, and tools for visualizing the results are necessary for researchers to glean insights from them.

See the 7 papers in this thread

Applied Bayesian nonparametrics

Bayesian nonparametrics provides powerful tools for designing flexible models whose complexity is allowed to grow with the amount of data. The observed data then automatically regularizes the model’s complexity, an elegant solution to the model selection conundrum. I have worked on spatial nonparametric priors and robust inference schemes for automatically segmenting images and videos, and on methods for discovering the parts of 3D object representations.

See the 6 papers in this thread

Federated learning and model fusion

Nonparametric methods for combining separately trained neural networks by matching their parameters.

See the 2 papers in this thread