In this theme we look at optimization and learning in directed and undirected graphical models. Undirected graphical models have factors that depend on many variables. This includes higher-order models as well as models with a large, even continuous, label space. Optimizing in such models is often NP-hard. Furthermore, it is often difficult to hand-craft the functional relationship between variables, hence it is necessary to learn them. The goal in this theme is to analyse the trade-offs between models, optimization and learning with the ultimate goal of achieving practically relevant algorithms which are efficient and accurate.
A shortlist of research topics we are excited about:
- optimization in undirected graphical models with higher-order factors, continuous label space, and models of very large size;
- combining generative and discriminative models;
- probabilistic learning and inference in undirected graphical models;
- combining deep directed models with undirected graphical models.
Running Projects:
- Diverse Solutions to Graphical Models
- Parallelizable Approximative Inference for Dense Non-Gaussian Graphical Models
- Exact MAP-Inference by Confining Combinatorial Search with LP Relaxation
- A link to the Open GM 2 project where we have been involved in
Recent (selected) publications and preprints:
Text-book: B. Savchynskyy Discrete Graphical Models – An Optimization Perspective Under review.
- S. Tourani, A. Shekhovtsov, C. Rother, B.Savchynskyy, MPLP++: Fast, Parallel Dual Block-Coordinate Ascent for Dense Graphical Models, ECCV 2018
- S. Haller, P. Swoboda and B. Savchynskyy. Exact MAP-Inference by Confining Combinatorial Search with LP Relaxation. AAAI 2018
- A. Shekhovtsov, P. Swoboda and B. Savchynskyy. Maximum Persistency via Iterative Relaxed Inference with Graphical Models. PAMI 2017.
- F. Michel, A. Kirillov, E. Brachmann, A. Krull, S. Gumhold, B. Savchynskyy, C. Rother. Global Hypothesis Generation for 6D Object Pose Estimation. CVPR 2017
- P. Swoboda, J. Kuske, B. Savchynskyy. A Dual Ascent Framework for Lagrangean Decomposition of Combinatorial Problems. CVPR 2017.
- A. Kirillov, A. Shekhovtsov, C. Rother, B. Savchynskyy. Joint M-Best-Diverse Labelings as a Parametric Submodular Minimization. NIPS 2016.
- Jörg H. Kappes, Bjoern Andres, Fred A. Hamprecht, Christoph Schnörr, Sebastian Nowozin, Dhruv Batra, Sungwoong Kim, Thorben Kroeger, Bernhard X. Kausler, Jan Lellmann, Nikos Komodakis, Bogdan Savchynskyy, Carsten Rother. A Comparative Study of Modern Inference Techniques for Discrete Energy Minimization Problems. IJCV 2015.
- P. Swoboda, B. Savchynskyy, J. Kappes, C. Schnörr. Partial Optimality by Pruning for MAP-inference with General Graphical Models, CVPR-2014 (Best Student Paper Award).
- V. Kolmogorov and C. Rother, Minimizing non-submodular functions with graph cuts – a review, PAMI,vol. 29, no. 7, 2007.