Research & Publications

Taxonomizing and Measuring Represenational Harms: A Look at Image Tagging

In this work, we examine computational approaches for measuring the "fairness" of image tagging systems, finding that they cluster into five distinct groups, each with their own analytic foundations. We further identify a diverse range of normative concerns that are often collapsed under the terms "unfairness," "bias," or even "discrimination" when discussing problematic cases of image tagging. In particular, we offer a four-part taxonomy of the harms that may be caused by image tagging systems, offering concrete examples of each. We then consider how different computational measurement approaches map to each of these harms. In so doing, we demonstrate that a wide range of approaches can help measure a wide range of harms. Our findings also highlight that no single measurement approach will provide a definitive result, nor will it be possible to infer from the chosen approach which harm researchers are actually seeking to measure. Lastly, equipped with a more granular understanding of the harms that may be cause by image tagging systems, we show that attempts to mitigate some of these harms may be in tension with one another.

Published and presented at the Thirty-Seventh Conference on Artificial Intelligence (AAAI 2023).

Shape the Vote

Our goal is to provide alternative tools to discuss election forecasting whereby voters are instead reminded and empowered to play their part in democracy.

Published an interactive data visualization in D3.js using multi-level regressions to allow voters to modify forecasted vote share to explore the possible outcomes of an election.

Detect suspicious IP addresses with the Amazon SageMaker IP Insights algorithm

IP Insights is an unsupervised learning algorithm for detecting anomalous behavior and usage patterns of IP addresses. IP Insights is natively integrated into the Amazon SageMaker platform. Therefore, all you need is to put your data in S3 and you can spin up fully distributed training cluster to train an IP Insights model and then deploy with one-click to an EC2 instance for real-time inference.

Get started with detecting suspicious IP addresses with the Amazon SageMaker IP Insights algorithm here.

DeepSurv: Personalized Treatment Recommender System Using A Cox Proportional Hazards Deep Neural Network

Researched the application of deep neural networks to survival analysis and demonstrated a deep neural network’s ability to predict the risk of an event occurring (i.e. death of a patient). Demonstrated state-of-the-art performance in predicting a patient’s risk of death and providing them with a personalized treatment recommendation. Released an open-source Python package with a Docker framework to increase the reproducibility of experiments.

Presented at the International Conference of Machine Learning Computational Biology Workshop 2016. Published in the journal BMC Medical Research Methodology.


Ran a social-media listening campaign to gather data on trans health issues. Analyzed over 12,000 responses and presented findings as policy recommendations to the U.S. Federal Government.

Presented as part of White House LGBTQ Tech & Innovation Summit 2015.