Varun Babbar

Varun Babbar

PhD student in Machine Learning @ Duke University

About

Hi! I'm Varun Babbar, a PhD student in Machine Learning at Duke University, advised by Cynthia Rudin. Previously, I worked as an ML Researcher at JP Morgan London, where I developed machine learning methods to accelerate software development. This included methods to generate embeddings of repositories, designing systems based on large language models to automate code development, and code unit test quality assessment based on weakly supervised learning (no links here because these were internal projects).

I graduated from the University of Cambridge in 2022 with a BA + MEng in Information and Computer Engineering, ranking in the top 5% of my class. During my time at Cambridge, I've collaborated with Umang Bhatt and Adrian Weller on human-aligned uncertainty quantification and risk control, Sean Moran on federated learning for Covid-19 diagnosis, and Rafal Mantiuk on developing visual loss functions for image to image translation models.

Research Interests

My research interests lie broadly in trustworthy machine learning and human-AI collaboration. In particular, I design algorithms, techniques, and frameworks that enable users to better understand the predictive models they are deploying as well as the data they train the models on. I like to think of my research as a tree:

My Research Tree

  1. The characteristics of test and training datasets must be fully understood. In particular, having differing training and test distributions of data risks deployment of a model that is flawed. In these cases, explaining distribution shifts using interpretable methods is more important than viewing datasets through the lens of basic statistical measurements.
  2. The predictive model must be interpretable, computationally inexpensive to run, and have some guarantees on optimality. The end user and all relevant stakeholders should fully understand the operation and the limitations of the model.
  3. When full interpretability is not possible (e.g. with heavily parameterized models such as LLMs), can we obtain sound explanations from the model?

Random Facts

Here are the cities I've lived in :
New Delhi     Mumbai     Singapore     Mumbai     Cambridge     London     Durham, North Carolina


Here are the languages I am fluent in: English, Hindi, French.


My non-sporting interests are quite random: Rubiks cubes, chess, and playing the ukulele!

I love playing basketball and squash: I was part of Cambridge's Varsity squash team. We toured a lot in and around Cambridgeshire, playing against county clubs and university teams alike. I'm still active in Duke's squash and basketball communities, though I only play for recreation.