Although machine learning has made a significant progress in recent years, there are still many open challenges, for example regarding the robustness, reliability and generality of learning methods. In order to further push the possibilities of learning system performance, in my work I aim to combine two key concepts, namely, (i) a deeper integration of of human knowledge in the form of expert-designed models and algorithms into learning methods, and (ii) exploiting very large-scale datasets for training. For both aspects, an important fundamental aspect is how to establish correspondences across the data items (e.g. images, or shapes). In this talk several approaches for alignment, matching and correspondence problems will be discussed, including combinatorial approaches for shape-to-image matching, convex relaxations for quadratic assignment problems, and the synchronisation of transformations when aligning multiple objects.
|Date||February 4, 2019 (Mon) 14:00 - 15:30|