Assistant Professor, Computer Science
Stanford AI for Human Impact Lab
Stanford Artificial Intelligence Lab
Statistical Machine Learning Group
AI Safety @Stanford
ebrun at cs dot stanford dot edu
Publications Awards News Group Teaching twitter: emmabrunskill
Recent: Keynotes at Conference on Learning Theory (COLT 2019) and Uncertainty in Artificial Intelligence (UAI 2019). Slides for the UAI talk available here.
My mission is to create reinforcement learning systems that help people live better lives. To do so, my lab works on the full stack of technical questions inspired by our applications of RL to healthcare and education. The key challenge is that since reinforcement learning systems learn through experience, both gathering and leveraging data impacts people. This prompts questions about how to most efficiently gather that data, and how to best use data to support the complex objectives inherent in people facing industries, such as robustness, performance, fairness and safety (see our recent Science article for one effort in this direction). RL systems make interventions (decisions/actions) based on perception, and therefore we also partner with collaborators to design and iterate on the intervention and perceptual spaces that are most effective, and how the RL system can interact with a human-in-the-loop to create outcomes that most benefit people.
Recent focuses include:
Foundations of efficient reinforcement learning. A key challenge is to understand the limits of how an agent should balance exploration versus exploitation. We have proved (to our knowledge) the first probably approximately correct (PAC) results for transfer reinforcement learning (UAI 2013), concurrent multi-task reinforcement learning (AAAI 2015), and partially observable reinforcement learning (AISTATS 2016). Recently we have closed the question of how much data is required in tabular episodic RL with minimax PAC bounds (DLWB, ICML 2019), and provided the first (to our understanding) instance-dependent regret bounds for tabular episodic RL (ZB, ICML 2019).
What If reasoning for sequential decision making. There is an enormous opportunity to leverage the increasing amounts of data to improve decisions made in healthcare, education, maintenance, and many other applications. Doing so requires what if / counterfactual reasoning, to reason about the potential outcomes should different decisions be made. We have introduced new statistical estimators to direct minimize the error of the predicted performance of new strategies (ICML 2016), provided ways of comparing across alternate model predictions (LAS 2017), analyzed the challenges of using importance sampling-based approaches for selecting policies (UAI 2017, best paper) and created new methods that can exponentially reduce the variance of the resulting estimators (NIPS 2017).
Human-in-the-loop systems. Artificial intelligence has the potential to vastly amplify human intelligence and efficiency. We are working on systems to train crowdworkers using (machine) curated material generated by other crowdworkers (CHI 2016), and identify when to expand the system specification to include new content (AAAI 2017) or sensors. We are also interested in ensuring machine learning systems are well behaved (Arxiv 2017) with respect to their human users' intentions, also known as safe and fair machine learning..
Interested students: Unfortunately I am unable to respond to most emails about openings for internships, graduate and postdoctoral positions in my group. Admission decisions are made at the department level so I will not be able to respond about your likelihood of acceptance or possibility of working with me. If you are already enrolled at Stanford or have been admitted, please feel free to reach out if you're interested in discussing research opportunities in my group. I accept (already admitted Stanford) students in my research group every year.