Print this page to PDF to save a copy. The content is editable for light modifications. Elements with hide-on-print (like this one) won't show. Use browser print settings to customize the output till relevant sections fit on one page. Feel free to copy this style! The format is inspired by Shreyas' (much better) resume.

Atharva Sehgal

[email protected]
atharvas.net
linkedin.com/in/atharvas

Education

PhD in Computer Science, University of Texas Austin
Advisor: Swarat Chaudhuri. Focus on neuro-symbolic algorithms for visual reasoning and scientific discovery. Also in program synthesis, computer vision, and foundation models.
August 2021 - Present
B.S. in Computer Science, University of Illinois Urbana Champaign
Graduated with high honors. Minor in Linguistics. James Scholar.
August 2017 - May 2021

Experience

Caltech (Yue Lab)
Visiting Student Researcher
June 2024
Working with Yisong Yue on structured machine learning algorithms with a focus on AI4Science.
UT Austin (Trishul Lab)
Graduate Researcher
June 2021 - Present
Developing neuro-symbolic techniques for visual reasoning, program synthesis, and structured learning. Working on foundational models for visual reasoning and scientific discovery.
UIUC (Madhusudhan Parthasarathy's Group)
Undergraduate Researcher
August 2020 - May 2021
Developed a novel dataset of visual discrimination puzzles (VDPs). Used Python/PyTorch/Tensorflow to develop and test computer vision models to achieve state-of-the-art performance on VDPs. Utilized few shot classification models like a scene-graph generator (Mask RCNN backbone), an object detector (YOLO backbone), a VAE prototypical model, and a triplet-loss contrastive model. Paper accepted to IJCAI 2022.
UIUC (Sasa Misailovic's Group)
Undergraduate Researcher
February 2020 - December 2021
Formulated and engineered a compiler for efficient low-precision probabilistic programming in C++17. Developed experiments for the project and built a testbench measuring power usage, accuracy, and runtime on various platforms. Paper accepted to DAC 2021.
InMobi
Data Science Intern
May 2018 - August 2018
Implemented three features for conversion rate and click-through rate prediction models extensively used within InMobi. These features were based on agglomerative clustering of geospatial data and organic installs. Results and analysis used to improve CVR prediction models.

Publications

LaSR: Symbolic Regression with a Learned Concept Library NeurIPS 2024
Arya Grayeli*, Atharva Sehgal*, Omar Costilla-Reyes, Miles Cranmer, Swarat Chaudhuri
Neurosymbolic Grounding for Compositional World Models ICLR 2024
Atharva Sehgal, Arya Grayeli, Jennifer J. Sun, Swarat Chaudhuri
Neurosymbolic Programming for Science AI4Science @ NeurIPS 2022
Jennifer J. Sun*, Megan Tjandrasuwita*, Atharva Sehgal*, Armando Solar-Lezama, Swarat Chaudhuri, Yisong Yue, Omar Costilla-Reyes
Composing Neural and Symbolic Reasoning with an Application to Visual Discrimination IJCAI/ECAI 2022
Adithya Murali, Atharva Sehgal, Paul Krogmeier, P. Madhusudan
Statheros: A Compiler for Efficient Low-Precision Probabilistic Programming DAC 2021
Jacob Laurel, Rem Yang, Atharva Sehgal, Shubham Ugare, Sasa Misailovic

Projects

tacc-inference Library August 2024
The tacc-inference library provides a common API to run LLMs on a single node or multiple nodes of TACC's Vista and Frontera supercomputers (the largest academic supercomputer in the US). Used by 50+ labs. (pip install tacc-inf)
neurosym Library August 2023
The neurosym library is a Python package for neurosymbolic program synthesis. It is the first framework which integrates tools for DSL design, program search, and program abstraction in a self contained package. Used extensively in research, production, and teaching. Joint work with collaborators at MIT. (pip install neorosym)
Programmatic Structured Pruning of CNNs May 2022
Developed a tensor programming language to describe any CNN network and designed a novel synthesis mechanism to hierarchically distill a CNN into an executable program. Achieved a 98% compression ratio with only a 1% accuracy drop on the CoCo dataset.
EuclidTrainer December 2021
Utilized Euclidean geometry for calculating precise 3D pose estimations from 2D pose estimation models for static videos. Applied this to create a weight training recommendation algorithm.

Technical Strengths

Computer Languages: Python, C, Julia, C++14, Haskell, HTML/CSS/JavaScript, OCaml
Frameworks: PyTorch/TensorFlow/Scipy, Pandas/Dask, NetworkX, Coq/Lean, Z3, Pyro

Outreach, Service, and Talks

Academic Reviewing
ICML (2023, 2024), NeurIPS (2022, 2023, 2024), ICLR (2023, 2025), AISTATS (2025)
Talks
Neurosymbolic Programming and Scientific Discovery Chalmers University
Tutorial on Neurosymbolic Programming Neurosym Summer School (2022), POPL (2023), ICSE (2024), Neurosym Summer School (2024)
tinyurl.com/neurosym
Teaching
Math Tutor Caltech Y Youth of Promise (FA24)
College Math Prep (Co-Instructor) Coleman State Prison (FA23)
DiRP: Neurosymbolic Programming (Instructor) UT Austin (SP24, FA23, SP23)
Honors: Embedded Systems (TA) UIUC (SP20)
Honors: Algorithms for String Processing (TA) UIUC (FA20, SP21)
Data Structures and Algorithms (TA) UIUC (FA19, SP20, FA20, SP21)
Discrete Mathematics (TA) UIUC (FA20)

Relevant Coursework

Program Synthesis, Computer Vision, Robot Learning, Data Driven Algorithm Design, Programming Languages, Trustworthy ML