Recent work: Uncertainty Quantification: In machine learning we have focused on well calbibrated uncertainty quantification - that is algorithms that quantify the uncertainty in their predictions and do so well. This is important to downstream applications that use the predictions. Most of our work involves algorithms for approximate inference in Bayesian graphical models, including new formulations for approximate inference, theoretical analysis providing performance guarantees, optimization algorithms and implementations, and applications in active learning and robotics. See recent publications in NIPS 2017 , arxiv 2018 , AABI 2019 , AISTATS 2021 , ICBINB 22 , RSS 2022 .
Recent work: Planning and Inference: In planning/control/RL: our work has focused on scallable planning algorithms with an emphasis on the connection between planning and inference and the insights this connection provides. We have developed a new family of algorithms that have top performance in challenging environments, and have developed a framework comparing different approximate inference schemes for planning. Our algorithms operate in large combinatorial spaces as well as contunuous spaces, with applications in robotics. See recent papers in NeurIPS 2018 , NeurIPS 2019 , ICAPS 2019 , Lifted Inference Book 2021 , PGM 2022 , IJCAI 2023 .
Short overview talks: coming soon (see some links in publications page).
Visualizations from applications in robotics:
Planning through Approximate Inference
Contolling an Autonomous Ground Vehicle / Surface Vessel
|Uncertainty Quantification for Robotic Information Gathering
Additional videos and demonstrations available at: DiSProD Project Page
Additional videos and demonstrations available at: AK Project Page
Older Work: includes contributions in computational learning theory, logical reasoning in propositional and first order logic, learning in logical settings (inductive logic programming), planning in logical spaces (using novel data structure First Order Decision Diagrams), decision theoretic planning, frequent set mining. Early work introduced the ideas of Learning to Reason and Learning to Act as formal frameworks for analysing machine leanring systems which go beyond classification and prediction.
For more information please see: (see publications)