Project aimed at detecting and identifying different adversarial attacks on models trained for various natural language processing (NLP) tasks such as abuse detection, sentiment analysis, machine summarization, machine translation, and fake news detection. Funded by DARPA.
A new unlearning procedure for random forest models that enables them to be updated efficiently in response to deletion requests; on average, we found these new models to be anywhere from two to four orders of magnitude faster than retraining the model from scratch on a revised dataset. Funded by DARPA.
Project aimed at identifying the most influential training samples responsible for a given tree-ensemble (e.g. random forest, gradient boosted trees, etc.) prediction using a surrogate kernel model. Funded by DARPA.
A thorough overview of the literature about model interpretability via the training data, and efficient data deletion mechanisms for current machine learning models.
Large collaboration project across 25+ universities and institutions aimed at better detecting manipulated images and videos. Our role was to combine the outputs from smaller teams that analyze different aspects of the image / video. Funded by DARPA.
Structured learning problem that uses the rich relational structure of social network data to more effectively detect spammers and spam messages. Funded by the Army Research Office (ARO).
An exploratory data analysis of all school shootings from 1774 - 2016.
A survey of the latest techniques used to process large graphs in a distributed manner.
This project was designed for people with hearing issues, allowing them to "see" sound from their surroundings by overlaying a graphical representation of that sound on a pair of augmented reality glasses.