I am a first-year PhD student at UC Berkeley advised by Dan Klein and Dawn Song. I work on Machine Learning and Natural Language Processing as part of Berkeley NLP, the RISE Lab, and Berkeley AI Research (BAIR).
Before this, I did my undergrad at the University of Maryland, where I worked with Jordan Boyd-Graber. I spent most of 2019 working at the Allen Institute for AI with Matt Gardner and Sameer Singh.
Interpretability We look to open up the black box of machine learning by interpreting model predictions. We have analyzed the limitations of existing interpretation methods and designed new techniques for generating saliency maps. We facilitate research and adoption of interpretation methods through an open-source interpretation toolkit. Our current research probes the naturally emergent knowledge of pre-trained NLP models and studies when interpretations are useful in practice.
Robustness We study how an adversary or a distributional shift can affect model behavior. We have created an adversarial attack called Universal Adversarial Triggers that exposes egregious failures of state-of-the-art NLP systems. We have also used adversarial attacks to reveal counterintuitive model behavior on reduced and human-modified inputs. Our current research works to increase model robustness and studies new attack vectors such as model stealing and data poisoning.
Dataset Biases We investigate how models can use subtle dataset patterns (annotation artifacts) to achieve high test accuracy without truly understanding a task. We have analyzed techniques for discovering these artifacts and applied them to identify issues in existing datasets. We also study how overreliance on dataset biases can cause downstream failures such as susceptibility to adversarial attacks. Our current research creates new ways to more comprehensively evaluate models.