Harmon Bhasin

I am a research assistant at SecureBio and a research affiliate at MIT Media Lab where I work under Kevin Esvelt. My interests lie in the intersection of computational biology and biosecurity.

I did my undergrad in CS from UW-Madison, where I had the privilege of being mentored by Sushmita Roy, Junjie Hu, and Stav Atir. During my time there, I developed methods for analyzing single cell omic datasets (scRNA-seq & scATAC-seq), with orthogonal work in transformer interpretability. I had the pleasure of starting two startups and one student organization, all of which were cool learning experiences.

If you have any questions or would like to connect, please feel free to reach out.

news

Jul 01, 2024	Subreviewer for Machine Learning in Computational Biology 2024 (MLCB 2024).
May 12, 2024	Graduated from the University of Wisconsin-Madison, with a B.S. in Computer Science.
Mar 15, 2024	First author paper accepted in the Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL 2024).
Dec 01, 2023	Subreviewer for Research in Computational Molecular Biology (RECOMB 2024).
Nov 28, 2023	Delivered a talk at the RECOMB/ISCB Conference on Regulatory & Systems Genomics with DREAM Challenges (RSGDREAM 2023).

selected publications

NAACL

How does Multi-Task Training Affect Transformer In-Context Capabilities? Investigations with Function Classes

Harmon Bhasin, Timothy Ossowski, Yiqiao Zhong, and Junjie Hu

In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics, Apr 2024

Abs HTML Code

Large language models (LLM) have recently shown the extraordinary ability to perform unseen tasks based on few-shot examples provided as text, also known as in-context learning (ICL). While recent works have attempted to understand the mechanisms driving ICL, few have explored training strategies that incentivize these models to generalize to multiple tasks. Multi-task learning (MTL) for generalist models is a promising direction that offers transfer learning potential, enabling large parameterized models to be trained from simpler, related tasks. In this work, we investigate the combination of MTL with ICL to build models that efficiently learn tasks while being robust to out-of-distribution examples. We propose several effective curriculum learning strategies that allow ICL models to achieve higher data efficiency and more stable convergence. Our experiments reveal that ICL models can effectively learn difficult tasks by training on progressively harder tasks while mixing in prior tasks, denoted as mixed curriculum in this work.