I'm a postdoc at University of Tübingen supervised by Ulrike von Luxburg. My research is broadly focused on developing theory that builds towards more robust and trustworthy machine learning systems. I received my Bachelor of Science in Mathematics from MIT and my Phd from UCSD in 2023 where I was fortunate to be advised by Kamalika Chaudhuri and Sanjoy Dasgupta.
I've lately been very interested in studying how effective and reliable current explainability and interpretability tools are. Thus far, my approach has been to study older (but still widely applied tools) feature attribution methods. I'm especially pleased with our work on the popular method SHAP where we found some cute mathematics that shows a simple modification under which KernelSHAP enjoys some reasonably strong provable guarantees.
One of the first problems I worked on was in understanding "adversarial examples," which are small imperceptible changes made at test-time designed to cause misclassification. The approach we took was to study this problem in the context of much more old-fashioned machine learning techniques such as linear classifiers and non-parametrics. We found that some of the phenomena encountered in deep networks (such as differing sample complexities for learning robust classifiers rather than accurate ones) also occur in these simpler settings. We also found some simple ways to mitigate these issues including a modification under which non-parmetrics converge towards the optimally robust classifier.
I was introduced to online clustering by Michal Moshkovitz and immediately loved it for its simple setting and interesting algorithms. One of our main results was a simple efficient algorithm that nevertheless achieved an O(1) approximation to the optimal loss. Our main innovations were in efficiently adjusting to data sequences where the relevant distance scale rapidly changes.
Looking for my research papers? Visit the publications page for a complete list of preprints and published work.