I lead research projects focusing on enhancing trust in machine learning and AI systems. Particular research themes include:* Sensitivity and Robustness of ML methods: Developed fast and accurate approximations tomethods for quantifying sensitivity of machine learning models to perturbations in training dataand modeling assumptions. These works led to publications at premier machine learning conferences NeurIPS and AISTATS, as well as patents. * Uncertainty quantification: Developed approaches for quantifying uncertainties in the predictions of modern deep neural networks, including methods for reliable model selection and improved calibrationof Bayesian neural networks, disentangling learned representations in deep generative models,and federated learning of neural networks. These led to publications at NeurIPS, ICML, UAI, and JMLR, in addition to contributions to IBM products and the development of open source tools-- UQ360 (https://github.com/IBM/UQ360)

I work on developing statistical models for longitudinal and time series data.* Probabilistic Methods for Longitudinal Data: Developed models and methods for drawing inferencesfrom irregularly sampled temporal data arising in healthcare settings. Applied developedmodels for studying the progression of neurodegenerative diseases, including Huntington’s disease,Parkinson’s disease, and ALS. These works have been published in high impact venues including: Lancet Digital Health, Nature Computational Science, Movement Disorders, and Machine learning for healthcare (MLHC). These works have also led to several patents and two outstanding technical achievement awards.

Statistical Modeling: Designed probabilistic models for discovering topics from text collections, regions from images/videos and parts from 3D objects.Bayesian Nonparametrics: Developed models for spatio-temporally correlated data that increase in complexity with increasing amounts of data.Inference: Built effective, robust, stochastic search and MCMC based inference algorithms for latent variable models.

High dimensional classification: Explored dictionary learning algorithms for learning efficient representations for classification of high dimensional continuous data (images). Large Scale Density Modeling: Developed parallel Expectation Maximization algorithmsfor learning large scale mixtures of factor analyzers from several million high dimensional data points.

Unsupervised Learning: Developed statistical models and Metropolis Hastings (MCMC) samplers for discovering the number and extent of regions exhibiting coherent appearance and motion in video sequences.

Document modeling: Utilized topic models for measuring the relevancy of user contributed documents to subject- themed digital libraries.Autonomous robot navigation: Designed computer vision algorithms and features for obstacle avoidance and traversable path identification to aid robots navigate in unstructured environments.

Learning from class proportions: Developed regularized EM algorithms for fitting finite mixture models utilizing non traditional sources of information such as prior knowledge of class proportions. These algorithms require only a fraction of training data and are more noise tolerant.

Geomorphic mapping: Created tools for categorizing planetary surfaces into geomorphic landforms (craters, ridges, etc.) using naive Bayes, SVM and bagged decision tree classifiers.

Telescope analytics: Identified and captured leading indicators of radio telescope failures. Developed an extensible system for periodic logging of the relevant factors in a MySQL database.