doc. Ing. Tomáš Werner, Ph.D.

Optimization is ubiquitous in nowadays computer vision and machine learning, examples being energy minimization in computer vision, inference in probabilistic graphical models, or training of classiťrers (such as SVM or deep neural networks). A recent feature ofthese tasks is their size - they can easily have millions ofvariables and (sparse) constraints. Due to this, classical optimization theory and algorithms are often inadequate. For instance, popular simplex and interior point methods are inapplicable to linear programs ofthat size already for their superJinear space complexity. Therefore, large-scale distributed optimization algorithms muts be used, such as coordinate minimization minimization, alternating direction method of multipliers (ADMM) or stochastic gradient descent. The aim of the proposed doctoral thesis is to develop novel theory and algorithms for large- scale optimization in computer vision and machine leaming, with emphasís on coordinate minimization.

In machine learning and computer vision, structural statistical models of images describe collections of images with certain characteristics in a concise way, which then facilitate tasks like image segmentation, scene reconstruction, and object detection/recognition. Two classes of such models have been widely successful: graphical models and pictorial structures. The student should design and analyze new models within these two classes (possibly combining them) and algorithms for inference and learning. Publications on prestigious international conferences are expected. http://cmp.felk.cvut.cz/~werner/