My research in Foundational Machine Learning seeks to identify the structure underlying learning: how it emerges in optimization dynamics, what models remember, which data matters, and how people and AI can learn to work together.
Before Purdue, I was a visiting researcher at Google Research and a postdoctoral scholar at UC Berkeley. I received my PhD from the University of Texas at Austin and my MTech from IIT Bombay. I have multiple PhD positions available.
Fall 2026 Postdoc positions at IPAI. If you are interested in working with me, please send me a note before you apply.
Ongoing research
The group studies fundamental questions about how learning algorithms use data, what they retain, and how they interact with people. These five directions organize our current work.
We study why simple greedy algorithms can perform far better than worst-case theory predicts by identifying structure in the spectral landscape. For determinant maximization, stable-rank windows create broad success valleys, while sharp spectral drops produce unavoidable failure cliffs. This view connects matrix spectra to practical subset selection and gives stronger guarantees for random features, near-identity kernels, and spiked-plus-noise models.
We study how optimization algorithms behave near different classes of minima, and how data coherence, batch size, and perturbation scale determine which solutions remain stable. Our current work develops a unified view of SGD and sharpness-aware methods, connecting local loss geometry to generalization, low-complexity representations, and the emergence of simplicity bias.
We investigate what it means for a trained model to remember an individual example and how that information appears in representations, loss geometry, and membership-privacy risk. We use these signals to design principled machine-unlearning methods that remove designated data while preserving useful structure learned from the retain set.
We design coresets, pruning methods, and spectral sparsifiers that make learning possible with substantially less data or computation. Rather than preserving only first-order statistics, our methods aim to retain gradients, curvature, and the shape of the full-data loss landscape, producing compact summaries that remain reliable under noise and distributional complexity.
We study when AI systems should align with human judgment, when they should complement it, and how they can adapt between those roles. This includes learning from subjective annotations, estimating collective perspectives, and building human-centered ensembles that route decisions according to the relative strengths of people and models.
