I’m interested in the conceptual foundations of machine learning and the ethical, epistemic, and social implications of AI systems. I also use computational models to study opinion dynamics, group decision-making, collaboration and wisdom of crowds. Below are some of my recent and ongoing projects:
What Do We Really Want from Interpretability?
Interpretability is one of the most discussed topics in machine learning today, but what additional information do we actually need from a model, beyond its predictions? The answer depends on our goals. My research explores the conceptual landscape of interpretability, with several ongoing projects on counterfactual explanations, accountability, diagnostic models, and internalist guarantees for formal reasoning.
SSHRC Insight Development Grant
Title: Beyond the Accuracy-Interpretability Tradeoff: Optimizing Human-AI Collaborations
(With Chris Smeenk)
This project investigates what kinds of information, beyond predictions, can enhance human-AI team performance. We also draw lessons from case studies in physics to explore how we might establish the reliability of models we don’t fully understand.
Can Social Media Unpolarize Us?
While it is often believed that polarization dominates public discourse, some evidence suggests that society is not as polarized as it appears. There’s also a puzzling correlation between opinions on unrelated topics (e.g., anti-climate action and pro-gun stances), suggesting that our multi-dimensional opinion space has collapsed into just two perceived dimensions.
I am interested in how misperceptions of opinion space contribute to polarization, and whether dynamically configuring social connections could help restore a multi-dimensional opinion space.