Hi! I am a Predoctoral Research Fellow at Microsoft Research (MSR), where I am working with Dr. Amit Sharma and Dr. Emre Kiciman. My research focuses on causal representation learning for Out-of-Distribution generalization. I am broadly interested in developing machine learning systems that are robust to data distribution shifts in real-world deployments.
I graduated from BITS Pilani, Pilani with a Bachelor’s in Computer Science in 2021. I was fortunate to pursue my undergraduate thesis at the MultiComp Lab in the Language Technologies Institute of Carnegie Mellon University, supervised by Prof. Louis-Philippe Morency and mentored by Paul Pu Liang and Yiding Jiang. My thesis was focused on multimodal reinforcement learning, specifically exploring how language grounding can assist in capturing affordance of objects and accelerate learning of autonomous agents.
Prior to MSR, I was a research intern at Adobe’s Media and Data Science Research (MDSR) lab, working on knowledge enhancement of language models to make robust factual and commonsense reasoning-aware predictions. I have also had the good fortune of working with Prof. Dr. Chris Biemann at Language Technology Lab, Universität Hamburg on semantic parsing for knowledge graph question answering. In the past, I have interned as a software engineer at Microsoft and worked on computer vision problems at MapmyIndia.
Way back, I was obsessed with being an astronaut (maybe someday?) and my fascination with space led me to joining Team Anant, a group of passionate undergraduate students building BITS Pilani’s first nanosatellite, where I had a great time designing control algorithms and brainstorming with some amazing people. I love running, playing basketball, and am trying to find time to revisit my passion to write.
Bachelor of Engineering in Computer Science, 2021
Birla Institute of Technology and Science (BITS) Pilani, Pilani
[Jan 2023] Our work discussing the necessity of modeling the data-generating process for OOD generalization accepted as notable-top-25% for oral presentation at ICLR 2023!
[Jul 2022] I will be attending ICML 2022 in Baltimore!
[Jun 2022] Paper accepted as Spotlight to Workshop on Spurious Correlations, Invariance, and Stability at ICML 2022.
[Apr 2022] Two papers from internship at Adobe accepted to NAACL 2022!
[Mar 2022] Short paper studying SPARQL semantic parsing baselines accepted to SIGIR 2022.
[Sep 2021] Joined Microsoft Research as a Research Fellow where I am working with Dr. Amit Sharma and Dr. Emre Kiciman!
[Sep 2021] Two papers accepted to Workshop on Commonsense Reasoning and Knowledge Bases (CSKB) at AKBC 2021.
[Aug 2021] Received Prof. V S Rao Foundation Best All-Rounder Award 2021!
[May 2021] Started as a research intern at Adobe’s Media and Data Science Research (MDSR) lab.
[May 2021] Attending ICLR'21!
[Mar 2021] Paper accepted to ICLR 2021 Workshop on Embodied Multimodal Learning.
[Jan 2021] Started my bachelor thesis at MultiComp Lab, CMU!
Worked on improving agent exploration in sparse reward environments by formulating structured intrinsic rewards. Devised a novel form of curiosity leveraging grounded question answering to encourage the agent to ask questions about the environment and be curious when the answers to these questions change.
Developed a web application for real-time hospital resource monitoring (beds, ICUs, ventilators). Integrated a mask detection model to provide real-time information regarding the percentage of people wearing masks at any location using live video feed.
Developed a fully functional compiler from scratch (in C) capable of lexical analysis, syntax tree creation, semantic analysis, static and dynamic type checking and generating executable assembly code. The artificial language supported constructs like dynamic memory allocation, loops, if-else ladders, switch statements, nested scopes and function calls.
Worked on knowledge enhancement of language models (LMs) by augmenting structured knowledge externally. Created a new masked pre-training corpus using Wikipedia hyperlinks to identify entity spans; trained LMs to retrieve contextually relevant knowledge via masked language modeling on this modified corpus.