Kaustubh S. Ponkshe

Researcher | AI Enthusiast | IIT Bombay Alumnus

Kaustubh_headshot.jpg

Mohamed bin Zayed University of Artificial Intelligence

Abu Dhabi, UAE

Email: kaustubhponkshe@iitb.ac.in

About Me

I am Kaustubh Ponkshe, a researcher passionate about democratizing AI through trustworthy, efficient, and scalable machine learning systems. I hold a Bachelor’s degree in Electrical Engineering and a Master’s degree in Artificial Intelligence from IIT Bombay, where I was awarded the Undergraduate Research Award for one of the best bachelor’s theses. I am currently working as a researcher with Prof. Praneeth Vepakomma, as a part of the CERT-Lab at MBZUAI.

My research focuses on addressing challenges in fine-tuning, federated learning, and privacy-preserving AI. Some of my notable projects include:

  • LoRA-SB: Achieved up to 90x parameter efficiency while maintaining competitive performance in fine-tuning.
  • Power-Learning: Developed a privacy-preserving framework for collaborative learning with formal differential privacy guarantees.
  • StructFormer: Proposed a novel structure-aware language model for document comprehension.

Feel free to explore my publications and projects.

Affiliations

  • Researcher at Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi
  • Former Research Intern at MIT Media Lab
  • Former Graduate Student at IIT Bombay

Thank you for visiting my page!

news

Jan 23, 2025 GuideQ accepted at NAACL Findings :sparkles:
Dec 15, 2024 StructFormer accepted at the AAAI Workshop on Document Intelligence! :sparkles:
Dec 10, 2024 Attending NeurIPS 2024 to present our work on Federated LoRA
Dec 01, 2024 Our latest work LoRA-Silver Bullet is out on arxiv 😄

selected publications

  1. Publication Preview
    Initialization Using Update Approximation for Low-Rank Fine-Tuning
    Kaustubh Ponkshe*, Raghav Singhal*, Eduard Gorbunov, and 3 more authors
    2025
    Under Preparation @ ICML 2025
  2. Publication Preview
    FedEx-LoRA: Exact Aggregation for Federated and Efficient Fine-Tuning of Foundation Models
    Raghav Singhal*, Kaustubh Ponkshe*, and Praneeth Vepakomma
    Oct 2024
    Poster, FITML Workshop @ NeurIPS 2024, Under Review @ ICLR 2025
  3. Publication Preview
    Power Learning: Privacy-preserving Embeddings
    Kaustubh Ponkshe, Praneeth Vepakomma, and Peter Kairouz
    Oct 2024
    Oral, Trustworthy ML Workshop @ CIKM 2024