Sahil Verma

I'm a PhD Student in the Department of Computer Science at the University of Washington in Seatte.

I'm interested in ML Interpretability and Robustness research. I'm advised by Jeff Bilmes and Chirag Shah. I graduated from Indian Institute of Technology Kanpur (IIT Kanpur) in 2019 with a B.Tech in Electrical Engineering and a minor in Computer Science. I have been fortunate to work with several amazing researchers over the course of my career. I have interned as an Applied Scientist at Amazon and at Arthur AI. During my undergrad years, I also interned at ETH Zurich, MIT, and National University of Singapore (NUS).

Email / CV / Scholar / Twitter / LinkedIn / Fun

Experiences

[2019 - Present] PhD in Computer Science and Engineering, University of Washington, Seattle
[Summer, 2024] Research Intern, Deep Learning Group at Microsoft Research
[Summer, 2023] Applied Scientist Intern, Amazon
[Summer, 2022] Applied Scientist Intern, Amazon
[Summer, 2020-21] Research Scientist Intern, Arthur AI
[Summer, 2019] Research Intern, ETH Zurich
[Summer, 2018] Research Intern, MIT
[Summer, 2017] Research Intern, National University of Singapore
[2015 - 2019] B.Tech in Electrical Engineering, IIT Kanpur

If you have any questions or want to collaborate, please reach out via email or Twitter! Always happy to chat.

Research

I'm interested in understanding how a ML model makes the decision it produces, specifically depending on the data it is trained on. I'm also interested in improving the robustness of ML models based on our understanding of how they make decisions. Currently, I'm interested in AI safety and security research. This area would be the biggest beneficiary of ML robustness research.

	How Many Van Goghs Does It Take to Van Gogh? Finding the Imitation Threshold Sahil Verma, Royi Rassin, Arnav Das, Gantavya Bhatt, Preethi Seshadri, Chirag Shah, Jeff Bilmes, Hannaneh Hajishirzi, Yanai Elazar We ask "how many images of a concept does a text-to-image model need for its imitation?". We posit this question as a new problem: Finding the Imitation Threshold (FIT) and propose an efficient approach (MIMETIC²) that estimates the imitation threshold without incurring the colossal cost of training multiple models from scratch.
	Effective Backdoor Mitigation Depends on the Pre-training Objective Sahil Verma, Gantavya Bhatt, Avi Schwarzschild, Soumye Singhal, Arnav Das, Chirag Shah, John P Dickerson, Jeff Bilmes Best Oral Paper Award at BUGS Workshop at NeurIPS 2023 We demonstrate that the effectiveness of backdoor removal techniques is highly dependent on the pre-training objective of the model. Based on the empirical findings, we suggest practitioners to train multimodal models using the simple constrastive loss owing to its amenability to be cleaned of backdoors.
	RecRec: Algorithmic Recourse for Recommender Systems Sahil Verma, Ashudeep Singh, Varich Boonsanong, John P Dickerson, Chirag Shah We propose RecRec that provides recourse for content creators of a recommender system, an often ignored community in the recommender system interpretability research.
	Post-Hoc Attribute-Based Explanations for Recommender Systems Sahil Verma, Chirag Shah, John P. Dickerson, Anurag Beniwal, Narayanan Sadagopan, Arjun Seshadri Best Student Paper Award at TEA Workshop at NeurIPS 2022 We propose RecXplainer that provide attribute based explanations for recommendations.
	Amortized Generation of Sequential Counterfactual Explanations for Black-box Models Sahil Verma, Keegan Hines, John P Dickerson We propose FastAR, a novel stochastic-control-based approach that generates sequential recourses. FastAR is model agnostic and black box.
	Counterfactual Explanations and Algorithmic Recourses for Machine Learning: A Review Sahil Verma, Varich Boonsanong, Minh Hoang, Keegan Hines, John P. Dickerson, Chirag Shah Best Paper Award at the ML-RSA Workshop at NeurIPS 2020 We survey close to 200 recently published papers about counterfactual explanations in ML.
	Removing biased data to improve fairness and accuracy Sahil Verma, Michael Ernst, Rene Just We propose a black-box approach to identify and remove biased training data to improve model fairness.
	Fairness Definitions Explained Sahil Verma and Julia Rubin We consolidate, define, and delineate over a dozen ML fairness definitions.

Teaching Experience and Community Services

	Teaching Assitant, Introduction to Software Engineering (CSE403), Fall 2023 Teaching Assitant, Introduction to Machine Learning (CSE416), Fall 2021
	Student Volunteer at ACM ESEC/FSE 2017, Paderborn, Germany Reviewer for EAAMO, XAIF, NeurIPS, AAAI, ICML, AIES, FAccT, IEEE Transactions on Artificial Intelligence, Data Mining and Knowledge Discovery, International Journal of Data Science and Analytics, and several workshops.

Template adapted from Jon Barron's website . Special thanks to our friendly AGI (in making) for assistance in customization.