Finale Doshi-Velez
Looking for my group? Check out our shiny group website.
Curious about my non-CS interests, including writing fiction? Check out my personal website.
Finally, my old PhD website has many older bits that haven't made it here.
Interests
How do we create AI support that not only helps people do what they need to do, but be who they want to be? At the Data to Actionable Knowledge (DtAK) lab at Harvard Computer Science, we recognize that AI support is almost always only one part of a larger task -- for example, an agent can help analyze data streams from your wearables and present you with statistical patterns that may provide health insights, but it cannot sleep or de-stress for you.
We develop methods in reinforcement learning, interpretability, and probabilistic methods to create agents that help people both in the immediate and longer term. Our work spans specific application domains (e.g. medicine, health and wellness, humanitarian crisis negotiation) as well as broader socio-technical questions around human-AI interaction, AI accountability, and responsible and effective AI regulation. Our work falls into three major technical areas:
- Decision-making under uncertainty (especially sequential decision-making): How can we optimize for the ultimate decisions of a human+AI team, rather than that of an RL agent alone? Relatedly, how can we provide useful information, even if we can't solve for a policy? How can we ensure that AI support on some narrow, proximal task (e.g. medication selection) helps in the context of the human's broader, distal task (e.g. improving health, having agency over one's health)? We also have had significant efforts around batch RL and off-policy evaluation to derive potential treatment policies from observational data.
- Interpretability and interaction: How can we expose key elements of a model or policy for expert inspection? How can we design for robustness toward both human and agent error? How can we take sophisticated requests from human users to rapidly personalize to their needs? How can we create agents that proactively identify useful information to share while also leaving the user engaged and in control?
- Probabilistic modeling and inference (especially Bayesian models): How can we characterize the uncertainty in large, heterogeneous data? How can we fit models that will be useful for downstream decision-making? How can we build models and inference techniques that will behave in expected and desired ways, especially as users request potential changes that are relevant to them?
Short Bio (+ CV, photo)
Finale Doshi-Velez is a Herchel Smith Professor in Computer Science at the Harvard Paulson School of Engineering and Applied Sciences. She completed her MSc from the University of Cambridge as a Marshall Scholar, her PhD from MIT, and her postdoc at Harvard Medical School. Her interests lie at the intersection of reinforcement learning, interpretability, and probabilistic methods to improve human+AI discovery and decision-making.
Selected Additional Shinies: BECA recipient, AFOSR YIP and NSF CAREER recipient; Sloan Fellow; IEEE AI Top 10 to Watch
Getting in touch
Office hours: Unfortunately, I don't have the capacity to hold general office hours this term. However, I'm almost always available before class (having lunch in our lab's common space) and can almost always chat a bit right after class (in the classroom).
Email: I get a lot of email, and my inbox is generally an absolute disaster. Unfortunately, many times I don't even have a chance to read all my email. Please help me by reading my website first, having an informative subject line, and pinging again if you feel like I might have lost your mail. Office hours don't require an email first, and are a great time to also catch me in person.
Phone: I don't have an office phone. Please email or stop by in person.