Paroma Varma

Ph. D. Student - Electrical Engineering - Stanford University

paroma [at] stanford [dot] edu - @paroma_varma

I am a fourth year Ph.D. student advised by Prof. Christopher Ré and affiliated with the DAWN, SAIL, and StatML groups. I am supported by the Stanford Graduate Fellowship and the National Science Foundation Graduate Research Fellowship.

My research interests revolve around weak supervision, or using high-level knowledge in the form of noisy labeling sources to efficiently label massive datasets required to train machine learning models. This includes systems like Snorkel and Snorkel MeTaL that allow users to write labeling functions rather than label data by hand. In this context, I’m also interested in using developer exhaust, byproducts of the data analytics pipeline, to simplify complex statistical and search-based problems.

Latest News

04/21/19 Our paper on Learning Dependency Structures for Weak Supervision Models accepted to ICML 2019!
04/10/19 Our workshop paper on weak supervision accepted to IEEE Intelligent Vehicles Symposium 2019!
03/22/19 Lab members use Snorkel MeTaL to achieve state-of-the-art performance on the GLUE Benchmark
10/15/18 Snuba: Automating Weak Supervision to Label Training Data accepted to VLDB 2019!


Learning Dependency Structures for Weak Supervision Models
ICML 2019

Fred Sala and I use a robust PCA-based algorithm to learn dependency structures among weak supervision sources without using any labeled data. We take advantage of the sparsity pattern in the structure and improve the sample complexity of existing efforts. We provide an information-theoretic lower bound on the minimum sample complexity of the weak supervision setting and empirically show that it improves over existing methods in terms of the quality of training labels generated. [pdf] [code]

Snuba: Automating Weak Supervision to Label Training Data
VLDB 2019

We explore how we can make weak supervision techniques easier to adopt by automating the process of generating noisy labeling heuristics. We introduce a system that takes as input a small, labeled dataset and a larger unlabeled dataset and assigns training labels to the latter automatically. It generates heuristics that each labels only the subset of the data it is accurate for, and iteratively repeats this process until the heuristics together label a large portion of the unlabeled data. We find that this method can outperform weak supervision with user-defined heuristics and crowdsourcing in many cases. [pdf] [code]

Babble Labble: Learning from Natural Language Explanations
ACL 2018, NeurIPS 2017 DEMO

Braden Hancock and I explore how we can use natural language explanations for why crowd workers provide the labels they do to label training data more efficiently. We automatically parse these explanations into executable functions and apply them to large amounts of unlabeled data. We find that collecting explanations allows us to build high quality training sets much faster than collecting labels alone. [pdf] [code] [blogpost] [demo video]

Efficient Model Search using Log Data

We present preliminary methods that use the logs generated while training complex deep learning models to predict the performance of models with different architectures. We find that without training any new models, we can predict how well a model architecture will perform according to different metrics and within training time constraints. [pdf]

Coral: Enriching Statistical Models with Static Analysis
NeurIPS 2017, NeurIPS ML4H 2017, MED-NeurIPS 2017

We introduce a weak supervision framework to efficiently label image and video training data given a small set of user-defined heuristics. We identify correlations among heuristics using static analysis and incorporate this information into a generative model that can optimally assign probabilistic labels to training data. We apply this method to video querying and medical image classification tasks, outperforming fully supervised models in some cases. [pdf] [blogpost] [video]

Socratic Learning: Finding Latent Subsets in Training Data
HILDA @ SIGMOD 2017, NeurIPS FILM 2016

We explore how we can find latent subsets in training data that affect the behavior of weak supervision sources. We automatically identify these subsets using disagreements between the discriminative and generative models and correct misspecified generative models accordingly. We improve upon existing relation extraction and sentiment analysis tasks and make these latent subsets interpretable for users. [pdf] [workshop] [blogpost] [video]



Snuba: Automating Weak Supervision to Label Training Data
Paroma Varma and Christopher Ré.
To appear at International Conference on Very Large Databases (VLDB), 2019

Weakly supervised classification of rare aortic valve malformations using unlabeled cardiac MRI sequences
Jason Fries, Paroma Varma, Vincent Chen, Ke Xiao, Heliodoro Tejeda, Saha Priyanka, Jared Dunnmon, Henry Chubb, Shiraz Maskatia, Madalina Fiterau, Scott Delp, Euan Ashley, Christopher Ré and James Priest.

Training Classifiers with Natural Language Explanations
Braden Hancock, Paroma Varma, Stephanie Wang, Percy Liang and Christopher Ré.
In Association for Computational Linguistics (ACL), 2018

Exploring the Utility of Developer Exhaust
Jian Zhang, Max Lam, Stephanie Wang, Paroma Varma, Luigi Nardi, Kunle Olukotun and Christopher Ré.
In Workshop on Data Management for End-to-End Machine Learning (DEEM) at SIGMOD, 2018


Inferring Generative Model Structure with Static Analysis
Paroma Varma, Bryan He, Payal Bajaj, Imon Banerjee, Nishith Khandwala, Daniel L. Rubin and Christopher Ré.
In Neural Information Processing Systems (NeurIPS), 2017

Automated Training Set Generation for Aortic Valve Classification
Vincent Chen, Paroma Varma, Madalina Fiterau, James Priest and Christopher Ré.
In Machine Learning for Health (ML4H), Neural Information Processing Systems (NeurIPS), 2017

Generating Training Labels for Cardiac Phase-Contrast MRI Images
Vincent Chen, Paroma Varma, Madalina Fiterau, James Priest and Christopher Ré.
In Medical Imaging meets NeurIPS (MED-NeurIPS), 2017

Augmenting Generative Models to Incorporate Latent Subsets in Training Data
Paroma Varma, Bryan He, Dan Iter, Peng Xu, Rose Yu, Christopher De Sa, Christopher Ré

Flipper: A Systematic Approach to Debugging Training Sets
Paroma Varma, Dan Iter, Christopher De Sa and Christopher Ré.
In Workshop on Human-In-the-Loop Data Analytics (HILDA) at SIGMOD, 2017


Socratic Learning
Paroma Varma, Rose Yu, Dan Iter, Christopher De Sa, Christopher Ré
In Future of Interactive Learning Machines Workshop (FILM), Neural Information Processing Systems (NeurIPS), 2016

Efficient 3D Deconvolution Microscopy with Proximal Algorithms
Paroma Varma, Gordon Wetzstein
In Computational Optical Sensing and Imaging, Imaging and Applied Optics, 2016

Nonlinear Optimization Algorithm for Partially Coherent Phase Retrieval and Source Recovery
Jingshan Zhong, Lei Tian, Paroma Varma, Laura Waller
In IEEE Transactions on Computational Imaging, 2016


Source Shape Estimation in Partially Coherent Phase Imaging with Defocused Intensity
Jingshan Zhong, Paroma Varma, Lei Tian, Laura Waller
In Computational Optical Sensing and Imaging, Imaging and Applied Optics, 2015

Design of a Domed LED Illuminator for High-Angle Computational Illumination
Zachary Phillips, Gautam Gunjala, Paroma Varma, Jingshan Zhong, Laura Waller
In Imaging Systems and Applications, 2015

In the Past

Previously, I worked on problems related to computational imaging. As an undergraduate at UC Berkeley, I studied phase retrieval via partial coherence illumination and digital holography in Prof. Laura Waller’s Computational Imaging Lab. I also rotated with Prof. Gordon Wetzstein’s Computational Imaging Group and looked at solving 3D deconvolution problems more efficiently.


At UC Berkeley, I was a teaching assistant for the first offering of EE16A: Designing Information Devices and Systems and helped develop course material for the class as well. I was also a teaching assistant for EE20: Structure and Interpretation of Signals and Systems.

This page was generated using Jekyll and uses CSS from Kevin Burke. Icons made by Freepik from is licensed by CC 3.0 BY