Research

Research Interests: Large Language Models, Multimodal Machine Learning, Embodied AI, and Reinforcement learning, Social and Ethical NLP.

• NeuLab and CLAW Lab - Carnegie Mellon University

  

  • Graduate Student Researcher [Aug 2022 - Present]

    • Advisors: Dr. Graham Neubig, Dr. Yonatan Bisk, PRIOR @ Allen Institute of Artificial Intelligence (AI2)
    • Research Area : Multimodal Machine Learning, Embodied AI, and Reinforcement learning
    • Implemented human-level performance of the AI2THOR agent in new environments in Virtual Reality (VR) using an Oculus headset - Includes replicating continuous movement to discretized steps in VR to human-arm movement in real-world, mapping force to Oculus grip, Opening and Closing objects in VR, and rendering procedurally generated randomized scenes in VR.
    • Worked on building reinforcement learning environments for the agent and implemented the Proximity Policy Optimization Algorithm (PPO) for these environments and agents. Working on integrating semantic maps which will help in improving the performance of the agent.

• MIDAS Lab, IIIT-Delhi

  

  • Research Intern [Aug 2018 - Jun 2019]

    • Advisors: Prof. Rajiv Ratn Shah, Prof. Ponnurangam Kumaraguru, Prof. Roger Zimmermann
    • Research Area : Natural Language Processing, Social-Media Analysis, AI for Social Good
    • Invented an LSTM cum transfer learning model for Hinglish offensive text classification over HEOT dataset.
    • Researched and trained own word embeddings for code-mixed languages. (Achieved state of the art accuracy: 87%)
    • Led the dataset curation, annotation and cleaning initiative in the lab for 5000+ hinglish tweets for model pipeline.
    • Worked on a couple of publications and presented the paper at AAAI-19 conference in Hawaii. [Poster][Video][Blog]

• Netaji Subhas Institute of Technology

  

  • Student Researcher [Feb 2019 - May 2019]

    • Advisors: Prof. M.P.S. Bhatia, Prof. Preeti Kaur
    • Research Area : Natural Language Processing, Embedding Optimization
    • Optimized MaltParser to create Hindi contextualized embeddings over a corpus of IIT-Bombay Hindi movie reviews.
    • Compared contextualized embeddings of different dimensions with vanilla counterparts like, Twitter Word2Vec, FastText Hindi, GloVe on Hindi sentiment analysis task (Observed nearly 15% improvement in accuracy). [code]