Genome-wide association studies (GWAS) have demonstrated that the vast majority of genetic variants associated with human traits are in the non-coding region of the human genome. Many studies have found that these variants regulate gene expression, often in specific cell types or tissues. Identifying genetic variants that have coordinated effects on multiple complex traits as well as gene expression (“pleiotropic effects”) have been made easier by recent large-scale efforts such as the GTEx project as well as the eQTL Catalogue developed by our group. Despite this, most studies have only performed pairwise analysis between specific traits and a small set of gene expression studies. While powerful, these pairwise analyses will miss the pleiotropic effects that genes and variants can have on a wide range of human traits. The aim of this PhD project is to integrate multiple large-scale datasets to identify high-confidence associations between genetic variants, genes and complex traits and to systematically study how the effects of genetic variants propagate from DNA to gene expression to higher level and more complex traits and how pleiotropy arises in these networks.
Automatically answering questions based on books and other narratives is challenging because it requires complex reasoning over long texts. One reason systems fail to answer questions in the narrative setting is the lack of implicit background knowledge that humans possess and habitually use. This kind of background information is often not explicitly expressed in the narrative. However, this information can be extracted from the common-sense knowledge bases such as ConceptNet, or utilized via large language models pretrained on texts containing common-sense information. This PhD project will develop methods for improving the generative neural question-answering systems in the long narrative setting by studying ways to incorporate common-sense background knowledge into the system. Using external background knowledge is expected to improve two critical parts in the long narrative setting - the retrieval component that retrieves the relevant parts of the narrative expected to contain the answer and the reasoning component that generates the response.
The proposed doctoral project aims to advance our knowledge on extremal geometric structure of Lipschitz-free spaces, and tensor products of Banach spaces by characterizing their Daugavet points and Δ-points.
A very recent research line of the well-known Daugavet property takes advantage of certain localisations. We say that a Banach space X has the Daugavet property if and only if the norm identity (the Daugavet equation) ‖ Id + T ‖ = 1 + ‖ T ‖ holds for every rank-1 (= for every compact) operator T: X → X, where Id denotes the identity operator. E.g. classical Banach spaces C[0,1] and L1[0,1] both have the Daugavet property. The Daugavet equation has proved useful in approximation theory and in the geometry of Banach spaces. Following the geometric characterisation to the Daugavet property, together with T. Abrahamsen and K. Pirk, the supervisors have defined Daugavet- and Δ-points as a natural extension to study the Daugavet property quantitatively. A Δ-point x of a Banach space is a norm-one element that is arbitrarily close to convex combinations of elements in the unit ball that are almost at distance 2 from x. If, in addition, every point in the unit ball is arbitrarily close to such convex combinations, x is a Daugavet point. A Banach space X has the Daugavet property if and only if every norm-one element is a Daugavet point. Therefore, the concept of Daugavet points and Δ-points can be faced as a local version of the Daugavet property.
One of our main goals is to characterise the Daugavet and Δ-points in tensor products between two Banach spaces. We think that this will help to settle the longstanding open problem whether the Daugavet property of X and Y implies the Daugavet property of projective or injective tensor products of X and Y.
Small and tiny object recognition is a challenge that has become more important due to rapid development in AI based security tool development. In this PhD we are focusing on recognition of tiny objects in 2D and 3D using single shot learning. We will investigate Multiview analysis for 3D objects to enable late fusion in the recognition pipeline. 3D objects will be presented in pointcloud form which will be produced through the project SilentBorder. The developed deep neural networks will be subject to network optimization to assure the developed solutions can be used in the real-time scenarios.
Objective of this project is to understand the processes that triggered and supported the emergence of multicellular life and to reconstruct the paleoenvironmental conditions and the atmospheric composition during the Ediacaran and Cambrian transition. This period of time witnessed fundamental changes in biogeochemical cycles with the final oxygenation of the atmosphere, global glaciation and warming events, large disturbances in global C-cycle and the onset of eukaryotic diversification and the emergence of animal life. The project aims to recognize the changes in weathering intensity and regime during the Ediacaran and Cambrian transition and tries to reveal its links to changes in atmosphere composition, ocean chemistry and multicellular life. In this project, unique and well-preserved sedimentary rock successions and continental weathering crusts (i.e. denudation surfaces) from different paleocontinents and basins across the Ediacaran and Cambrian transition.