Research interests and publications
I have the good fortune of being a research scientist at IBM Research where I am currently considering various algorithms and applications of (mostly stochastic) bilevel optimization in AI and machine learning and trying to theoretically understand the various properties of transformer based models such as in-context learning and compositional generalization. You can find more about this here, here and here. I have also studied the bilevel problem(s) appearing in Automated Machine Learning and AI (AutoML and AutoAI), developing various new (continuous and discrete) derivative-free algorithms and also understanding various theoretical aspects of the AutoML/AutoAI problem (more details here). I have also been lucky to be involved with the Lale project on “Gradual AutoML”, which allows one to succinctly craft sophisticated ML pipelines, and automate the search of these pipelines.
I also moonlight as a (ab)user of neuro-inspired algorithmic motifs, especially really high dimensional but really sparse latent representations, for novel “simpler” ways of learning (see here). I am also interested in the unsupervised problem of density estimation, both from a perspective of interpretability and computational efficiency. I have spent a good amount of time studying various all-pairs problems in computational geometry such as similarity search, range search, and density estimation, which have a naive quadratic computational complexity in the number of points, and we have explored algorithms that have both empirical and theoretical complexity as close as possible to linear. My thesis research was the related topic of similarity search, and how we can both (i) use fast similarity search to speed up ML, and (ii) how ML can be used to speed up similarity search.
Here is a (loose) categorization of my publications in the different areas:
You can find a chronologically sorted list of my publications on my Google Scholar profile.
Experience.