CORD19 Submission: Using Scibert-Embeddings augmented by MeSH words for COVID-19 literature clustering and retrieval.

Our submission to the CORD19 Kaggle Challenge embedding papers by extracting SciBert embeddings for Medical Subject Headings (MeSH) keywords. We employed our document embeddings for literature clustering and retrieval.

Submission

Code

Perscriptive Analytics: Transportation Logistics - 1st Place.

Our local search algorithm for Vehicle Routing placed 1st in our Prescriptive Analytics course. Our approach combined simulated annealing with evolutionary algorithms; at a high-level, we ran simulating annealing in parallel for the top-k feasible solutions, and, each time, decayed the initial temperatures using a hyperbolic function. We also clustered candidate problems to determine optimal hyperparameter configurations.

Slides

Code

Star Trek GPT-2

As a holiday gift for my dad, I fine-tuned HuggingFace's pretrained GPT-2 on all episode scripts from the original series.

Examining Regional Abortion Discourse in America Via Twitter

For David Kerter's graduate Anthropology Demography course, I collected over 200 thousand geo-tagged Twitter posts related to "Abortion," and clustered regions at the state and county level. My results indicate that Twitter discourses correlate with both public support for legalizing abortion at the state level and reproductive policy at the city level. These discourses appear to capture the circulation of tweets through ideological networks, and messages that articulate opinions through pro-life and pro-choice frames. Additionally, my analyses indicates that these data-driven approaches, while far from perfect, have the potential to augment existing demographic data, and aid advocacy groups by identifying regions with dissonance between grassroots sentiment and law.

Paper

Slides

NeurIPs Reproducibility Challenge Submission: "Sample-Efficient Deep Reinforcement Learning via Episodic Backward Update"

We participated in the NeurIPS 2019 Reproducibility Challenge. We attempted to replicate the results from Sample-Efficient Deep Reinforcement Learning via Episodic Backward Update (Lee et al. 2019), with modest success.

Submission

Code

OpenReview

Negotiating the Scientific: The Reconciliation of Top-Down and Bottom-Up Forces in the NSF),

The purpose of our paper is to chart how the National Science Foundation's practices are crafted from both the "bottom-up forces" of the scientific community and the "top-down" decision-making that originates from other parts of the state. Specifically, we will explore how conflicting interpretations and goals interface, and how they are ultimately are reconciled, then embedded in organizational policy. To understand these processes, we interviewed individuals ranging from Computer Science professors to former Assisant Directors of the NSF.

We also applied a data-driven lens to piece together how the NSF's grand awarding practices materialize in identifiable trends. For example, we analyzed frequent n-grams in NSF grant data. A sample visualization of some of our results from the CISE Directorate can be found here.

Ultimately, we found that ranking members of the organization are motivated by a deep sense of accountability to the scientific community, and interpret the advancement of knowledge as something innately in national interest. However, the NSF must contend with entities (e.g. Congress) that do not necessarily interpret a clear relationship between the two. Thus, the NSF must continually assure other parties of the the value in basic research and draft strategic priorities that align with the Executive Office. Through the language of "Broader Impact" and "Technical Merit" the foundation can simultaneously establish its role as a bastion of science, and as an influential organization that shapes society for the better. These terms also leave much to the interpretation of the scientific community, and provide breathing room within this complex political landscape.

Uber Developed: The Implications and Contestations of Ride Hailing Business in South Africa

Summary: In South Africa, Uber pivots towards the role of economic developer; the ride-hailing application promotes a narrative wherein it is the provider of both job opportunities and socio-economic mobility. However, development is not an apolitical activity. In South Africa, Uber engages with a transportation architecture designed to differentiate classes of South African citizen. Uber's interface with this constructed landscape, and attempt at extending it, accentuates systems of domination. Uber's platform continues to reinscribe the racialized disparities with wealthy and poor; it affords the former with a convenient means of interacting with the landscape, inaccessible to the latter, and provides a means for transfiguring the unemployed into a highly controlled "flexible" labor source. However, the relation between Uber and transportation infrastructure is not unidirectional. In turn, South Africa's reality imposes itself on the company, forcing to to adapt to an environment where it must contend with colonial associations, a corrupt police force, and overt violence.

Paper

Quantifying Ideology in the Rhode Island Senate

As Associate Director of Brown Political Review's Data Board, I wrote and oversaw data-driven projects and articles. In Quantifying Ideology: Data-driven philosophical scales, I applied dimensonality reduction to legislator voting records to tease out ideological clusters in the Rhode Island Senate.

Article

CS1400 AI TRON Compeition -2nd Place

My submission to the CS1400 TRON competition placed second out of 22 teams. The class environment was a modified version of the original TRON game, where players compete to wall off their opponent. My implementation appled MiniMax with Alpha-Beta pruning, and evaluated states with the Voronoi heuristic weighted towards power-ups. To handle stochastic power-up effects, I modified the vanilla MiniMax algorithm into Expecti-Max by running multiple expansions for each random state transition, anding compute the value of an action in expectation.

ACLU Tech for Libery: Mapping Stop-And-Frisk, Arrest Data in Massachusetts

As an intern with the Tech For Liberty Team, I analyzed Boston Police Department stop-and-frisk data and Massauchusetts low-level arrest records obtained by the ACLU of Massachusetts. Additionally, I built internal web tools to visualize these records, and collate them with external data sources (e.g. Census data). My work was part of a broader push by the ACLU to elucidate 'broken windows' policing tactics, which disproportionately target marginalized communities, and people of color.

Empor.co

Beginning in the summer of 2016, I co-founded a start-up with my close friend Lucas Rosenblatt. We started by building a mobile application for college students to buy and exchange goods and services. However, with the advent of Facebook buying-and-selling pages, we pivoted to a clothing recommender engine. Our application learned users' shoe preferences to make personalized recommendations. While Lucas and I ultimately disbanded our company, we are thankful for the invaluable (and humbling) opportunity to build a commericalizable product from scratch.