One of my favorite drawings by Soviet mathematician and artist, Anatoly Fomenko. It represents a deformation of the Riemann surface of an algebraic function. Surprisingly, the surface is homeomorphic to a two-dimensional torus!
I am a research associate working with Tom Goldstein at the University of Maryland on various topics in machine learning. Specifically, my projects include differentially private dataset release, defenses for data poisoning and backdoor attacks, explainability for vision networks with parameter saliency, and algorithm-performing neural networks. Before this position, I graduated from Caltech with a B.Sc. in mathematics and a focus on geometry and topology.
June 2022 - Present
UCHICAGO PhD CANDIDATE (Computer Science)
Cumulative GPA: N/A
September 2015 - June 2019
BACHELOR OF SCIENCE CALTECH (Mathematics)
Cumulative GPA: 3.6
Wasserman Scholar 2016-2019
August 2013 - June 2015
PART-TIME STUDENT JOHNS HOPKINS UNIVERSITY
Cumulative GPA: 3.5
Future Scholars in Mathematics program
August 2011 - May 2015
TOWSON HIGH SCHOOL
Cumulative GPA 5.96
UMD Research Associate
Prof. Tom Goldstein
June 2020 - June 2022
Showed data augmentations offer state of the art defense against data poisoning and backdoor attacks. Published and presented this work at ICASSP 2021.
Used a combinatorial trick to prove rigorous differential privacy guarantees for the combination of mixup augmentation and additive isotropic Laplacian noise. Connected this result with work from ICASSP 2021 via the theoretical bounds on data poisoning efficacy by Ma et al. (2019). A preliminary version of this project was published and presented in the ICLR 2021 workshop on Security and Safety in Machine Learning Systems. The full work is under review at CVPR 2022.
Collaborated on a project which utilized parameter gradients for CNN-based vision networks aggregated and normalized on a filter-by-filter basis to assess why models incorrectly classify test examples. i.e. Saliency maps where gradient is taken with respect to parameter space instead of input space. A preliminary version of this work was published in ICLR 2021 workshop on Responsible AI. The full work is under review at ICLR 2022.
Helped develop and code base and run experiments for a project which uses weight sharing in iterated residual blocks of a fully convolutional neural network to solve a benchmark suite of algorithmic problems (prefix sums, mazes, and chess puzzles). This work was published at Neurips 2021. Worked on a follow-up to this paper which uses a modified optimization protocol and carefully adds skip connections to solve the "overthinking" problem and drastically improve performance across all three benchmark tasks. This work is under review at ICLR 2022.
TRIUMPH MFG CO-FOUNDER
Aug 2020 - Oct 2021
Manufactured and distributed concrete weight plate molds during the Covid-19 induced steel weight plate shortage.
Researched vacuum forming techniques, created a design for the product using CAD software, fashioned a positive mold out of wood, and commissioned BK Plastics in Odessa, FL for mass manufacturing.
Created sole proprietorship business, obtained re-sale certificate, and created an amazon seller account for hosting, warehouse storage, and distribution. Sold approximately one thousand units over the course of a year.
Learned google search SEO to promote the company website, and paid for advertising using Instagram, Amazon, and Google.
Prof. Martin Nowak
Collaborated with a graduate student in the Nowak lab on a project in statistical oncology
Used techniques from Markov chains and network flow theory to model metastasizing tumor networks, which seed each other at potentially variable rates.
Jan 2020 - Mar 2020
Prof. Matilde Marcolli
Summers 2017 & 2018
Studied the Bruhat-Tits tree in the context of non-commutative geometry, utilizing the interplay between geometry and algebra in which one encodes geometric properties as algebraic structures, applies robust algebraic tools, and pulls back algebraic results to develop new geometric insights.
Learned to comb through existing literature in a complicated field to get a grasp on open questions and the common approaches used to tackle them
Worked closely on the project with a graduate student and post-doc, and reported relevant findings in weekly group meetings with professor
MATH 0 TA
Summers 2017 & 2018
Close-read student assignments on a diverse set of introductory modules, correcting logical flaws and emphasizing the appropriate use of syntax in proof-writing.
Led weekly office hours, where students used a chat interface to remotely ask questions about the lecture notes and homework assignments.
Followed forum threads to answer course questions and related curiosities posted by students.
JHMI REASEARCH ASSISTANT
Prof. Mario Bianchet
June 2013 - Sep 2015
Expressed and purified full-length MICAL in E. coli, a protein integral to the mechanism of axonal guidance which is notoriously difficult to express in bacteria due to its toxicity.
Independently designed and conducted of characterization experiments using the full-length MICAL yield.
Presented my research and exchanged feedback during weekly lab meetings.
Won Intel STS 2015 semifinalist and later published two papers as a result of my work.
THE JUDY-BENJAMIN PROBLEM
The Judy-Benjamin problem is a famous example used as an objection to initialization of priors via maximum entropy. I use an information field theoretic model to neatly resolve the confusions stemming from the problem.
LEFSCHETZ FIXED POINT THEOREM
I present an exposition of the Lefschetz fixed-point theorem and its consequences. This is the most powerful theorem for characterization of fixed points of functions from a finite simplicial complex to itself.