Research

Broadly, I am interested in questions of trust in a machine learning (ML) analysis. I work on both tools to detect lack-of-robustness in ML analysis and methods for “robustifying” ML analysis through carefully designed models, algorithms that produce well-calibrated uncertainties, and are robust to poor optima.

Robustness to perturbations in data or modeling assumptions

Will the inferences drawn from a particular analysis or predictions made by a model change substantially under perturbations to training data, minor variations of modeling assumptions, or upon using alternate learning and inference algorithms? In this line of research, we develop tools for answering these questions.

A naive approach to understanding the effect of data perturbations involves refitting the model of interest to many perturbations of the data. This is infeasible for large datasets and structured latent variable models, which involve expensive marginalization over latent variables. We develop efficient but accurate approximations which 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 can be used to approximate cross-validation in hidden Markov models and Markov random fields.

We can also consider the effect of modeling assumptions on inferences drawn from an ML analysis. We look at this question in the context of Gaussian processes and develop a methodology for measuring sensitivity to the choice of the kernel choice. 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.

Uncertainty quantification in neural networks

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. Bayesian neural networks (BNN) hold the promise of retaining their point-estimated counterparts’ predictive performance while providing well-calibrated uncertainties and principled approaches for model selection. These potential advantages have motivated my research into BNNs. My work has explored the effects of commonly used priors and approximate inference algorithms on the quality of posterior uncertainties in BNNs, and developed algorithms for inference and efficient decision making in BNNs.

I am also interested in uncertainty quantification more broadly. I am a core contributor to the Uncertainty quantification UQ360 — an open source toolbox that provides a number of approaches to quantifying, measuring the qualtiy, and communicating uncertainties. A white paper describing the toolbox:

Statistical models for healthcare: Hypothesis generation and disease progression Models

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

These representations are useful for characterizing the progression of diseases from longitudinal follow up of patients. Variants of hidden Markov models are effective for characterizing disease progression as a sequence of jumps between interpretable disease states.

Tools for visualizing the results from such progression models are necessary for researchers to glean insights from such progression models.

Applied Bayesian non-parametrics for computer vision and model fusion

Bayesian nonparametrics (BNP) provides powerful tools for designing flexible Bayesian models whose complexity is allowed to grow with the amount of data. Observed data thus automatically regularizes the model’s complexity and provides an elegant solution to the model selection conundrum. I have worked on developing spatial BNP (and BNP inspired) priors and robust inference schemes for automatically segmenting images and videos.

Methods for discovering parts of 3D object representations.

BNP based Methods for federated learning and model fusion.

See here for an up to date list of publications.