Refine
Year of publication
Document Type
- Master's Thesis (42) (remove)
Language
- English (42) (remove)
Has Fulltext
- yes (42)
Is part of the Bibliography
- no (42)
Keywords
- WaterGAP (2)
- global water model (2)
- (n (1)
- AI Safety (1)
- ALICE (1)
- Activism (1)
- App ecosystem (1)
- Autonomous Driving (1)
- Bayesian Inference (1)
- Beryllium-7 (1)
Institute
Goal-Conditioned Reinforcement Learning (GCRL) is a popular framework for training agents to solve multiple tasks in a single environment. It is cru- cial to train an agent on a diverse set of goals to ensure that it can learn to generalize to unseen downstream goals. Therefore, current algorithms try to learn to reach goals while simultaneously exploring the environment for new ones (Aubret et al., 2021; Mendonca et al., 2021). This creates a form of the prominent exploration-exploitation dilemma. To relieve the pres- sure of a single agent having to optimize for two competing objectives at once, this thesis proposes the novel algorithm family Goal-Conditioned Re- inforcement Learning with Prior Intrinsic Exploration (GC-π), which sep- arates exploration and goal learning into distinct phases. In the first ex- ploration phase, an intrinsically motivated agent explores the environment and collects a rich dataset of states and actions. This dataset is then used to learn a representation space, which acts as the distance metric for the goal- conditioned reward signal. In the final phase, a goal-conditioned policy is trained with the help of the representation space, and its training goals are randomly sampled from the dataset collected during the exploration phase. Multiple variations of these three phases have been extensively evaluated in the classic AntMaze MuJoCo environment (Nachum et al., 2018). The fi- nal results show that the proposed algorithms are able to fully explore the environment and solve all downstream goals while using every dimension of the state space for the goal space. This makes the approach more flexible compared to previous GCRL work, which only ever uses a small subset of the dimensions for the goals (S. Li et al., 2021a; Pong et al., 2020).
Analysis of machine learning prediction quality for automated subgroups within the MIMIC III dataset
(2023)
The motivation for this master’s thesis is to explore the potential of predictive data analytics in the field of medicine. For this, the MIMIC-III dataset offers an extensive foundation for the construction of prediction models, including Random Forest, XGBOOST, and deep learning networks. These models were implemented to forecast the mortality of 2,655 stroke patients.
The first part of the thesis involved conducting a comprehensive data analysis of the filtered MIMIC-III dataset.
Subsequently, the effectiveness and fairness of the predictive models were evaluated. Although the performance levels of the developed models did not match those reported in related research, their potential became evident. The results obtained demonstrated promising capabilities and highlighted the effectiveness of the applied methodologies. Moreover, the feature relevance within the XGBOOST model was examined to increase model explainability.
Finally, relevant subgroups were identified to perform a comparative analysis of the prediction performance across these subgroups. While this approach can be regarded as a valuable methodology, it was not possible to investigate underlying reasons for potential unfairness across clusters. Inside the test data, not enough instances remained per subgroup for further fairness or feature relevance analysis.
In conclusion, the implementation of an alternative use case with a higher patient count is recommended.
The code for this analysis is made available via a GitHub repository and includes a frontend to visualize the results.
The reanalysis products and derived products, ERA5 (Copernicus Climate Change Service, 2018) and W5E5 (WATCH Forcing Data (WFD) methodology applied to ERA5) (LANGE ET AL., 2021) have been recently published initiating a new phase of scientific research utilizing these datasets. ERA5 and W5E5 offer the possibility to reduce insecurities in model results through their improved quality compared to previous climate reanalyses (CUCCHI ET AL., 2020). The suitability of either climate forcing as input for the hydrological model WaterGAP and the influence of the models specific calibration routine has been evaluated with four model experiments. The model was validated by analysing the models ability to produce reasonable values for global water balance components and to reproduce observed discharge in 1427 basins as well as total water storage anomalies in 143 basins using well established efficiency metrics. Bias correction of W5E5 was found to lead to more global realistic mean precipitation and consequently discharge and AET values. In an uncalibrated model setup ERA5 results in better performances across all efficiency metrics. Model results produced with W5E5 as climate input were strongly improved through calibration ultimately leading to the best performances out of all four model experiments. However, model performances considerably improved through calibration with both climate forcings hence calibration was found to have the strongest effect on model performance. Furthermore, spatial differences in performance of either forcing were identified. Snow-dominated regions show an overall better performance with ERA5, while wetter and warmer regions are better represented with W5E5. Finally, it can be concluded that W5E5 should be preferred as climate input for impact modelling; however, depending on the spatial scale and region ERA5 should at least be considered, in particular for snow-dominated regions.
We give theorems about asymptotic normality of general additive functionals on patricia tries, derived from results on tries. These theorems are applied to show asymptotic normality of the distribution of random fringe trees in patricia tries. Formulas for asymptotic mean and variance are given. The proportion of fringe trees with 𝑘 keys is asymptotically, ignoring oscillations, given by (1−𝜌(𝑘))/(𝐻 +𝐽)𝑘(𝑘−1) with the source entropy 𝐻, an entropy-like constant 𝐽, that is 𝐻 in the binary case, and an exponentially decreasing function 𝜌(𝑘). Another application gives asymptotic normality of the independence number and the number of 𝑘-protected nodes.
AI-based computer vision systems play a crucial role in the environment perception for autonomous driving. Although the development of self-driving systems has been pursued for multiple decades, it is only recently that breakthroughs in Deep Neural Networks (DNNs) have led to their widespread application in perception pipelines, which are getting more and more sophisticated. However, with this rising trend comes the need for a systematic safety analysis to evaluate the DNN's behavior in difficult scenarios as well as to identify the various factors that cause misbehavior in such systems. This work aims to deliver a crucial contribution to the lacking literature on the systematic analysis of Performance Limiting Factors (PLFs) for DNNs by investigating the task of pedestrian detection in urban traffic from a monocular camera mounted on an autonomous vehicle. To investigate the common factors that lead to DNN misbehavior, six commonly used state-of-the-art object detection architectures and three detection tasks are studied using a new large-scale synthetic dataset and a smaller real-world dataset for pedestrian detection. The systematic analysis includes 17 factors from the literature and four novel factors that are introduced as part of this work. Each of the 21 factors is assessed based on its influence on the detection performance and whether it can be considered a Performance Limiting Factor (PLF). In order to support the evaluation of the detection performance, a novel and task-oriented Pedestrian Detection Safety Metric (PDSM) is introduced, which is specifically designed to aid in the identification of individual factors that contribute to DNN failure. This work further introduces a training approach for F1-Score maximization whose purpose is to ensure that the DNNs are assessed at their highest performance. Moreover, a new occlusion estimation model is introduced to replace the missing pedestrian occlusion annotations in the real-world dataset. Based on a qualitative analysis of the correlation graphs that visualize the correlation between the PLFs and the detection performance, this study identified 16 of the initial 21 factors as being PLFs for DNNs out of which the entropy, the occlusion ratio, the boundary edge strength, and the bounding box aspect ratio turned out to be most severely affecting the detection performance. The findings of this study highlight some of the most serious shortcomings of current DNNs and pave the way for future research to address these issues.
Statistical shape models learn to capture the most characteristic geometric variations of anatomical structures given samples from their population. Accordingly, shape models have become an essential tool for many medical applications and are used in, for example, shape generation, reconstruction, and classification tasks. However, established statistical shape models require precomputed dense correspondence between shapes, often lack robustness, and ignore the global surface topology. This thesis presents a novel neural flow-based shape model that does not require any precomputed correspondence. The proposed model relies on continuous flows of a neural ordinary differential equation to model shapes as deformations of a template. To increase the expressivity of the neural flow and disentangle global, low-frequency deformations from the generation of local, high- frequency details, we propose to apply a hierarchy of flows. We evaluate the performance of our model on two anatomical structures, liver, and distal femur. Our model outperforms state-of-the-art methods in providing an expressive and robust shape prior, as indicated by its generalization ability and specificity. More so, we demonstrate the effectiveness of our shape model on shape reconstruction tasks and find anatomically plausible solutions. Finally, we assess the quality of the emerging shape representation in an unsupervised setting and discriminate healthy from pathological shapes.
During RUN3 (2021-2023) of the Large Hadron Collider, the Time Projection Chamber (TPC) of ALICE will be operated with quadruple stacks of Gas Electron Multipliers (GEMs). This technology will allow to overcome the rate limitation due to the gated operation of the Multi-Wire Proportional Chambers (MWPCs) used in RUN1 (2009-2013) and RUN2 (2015-2018).
As part of the Upgrade project, long-term irradiation tests, so called "ageing tests", have been carried out. A test setup with a detector using a quadruple stack of 10x10cm2 GEMs was built and operated in Ar-CO2 and Ne-CO2-N2 gas mixtures. The detector performance such as gas gain and energy resolution were monitored continuously. In addition, outgassing tests of materials used for the assembly process of the upgraded TPC were performed. To reach the expected dose of the GEM-based TPC, the detector was operated at much higher gains than the TPC. It was found, that the GEMs could keep their performance within the projected lifetime of the TPC. Most of the tested materials showed no negative impact on the detector. For the tested epoxy adhesive no certain conclusion could be drawn.
At much higher doses than expected for the upgraded TPC, a new phenomenon was observed, which changed the hole geometry of the GEMs and led to a degradation of the energy resolution. Even though its occurrence is not expected during the lifetime of the GEM-based TPC, simulations were carried out to study this effect more systematically. The simulations confirmed, that a change of the hole geometries of the GEMs, lead to an increase of the local gain variation, which results in a decrease of the energy resolution.
Furthermore the effect of methane as quench gas on GEMs was studied, even though this gas is not foreseen to be used in the TPC. From ageing tests with single-wire proportional counters it is well known that hydrocarbons are produced in the plasma of the avalanches, which cover the electrodes and lead to a degradation of the detector performance. Even though GEMs have a quite different geometry, the ageing tests showed, that also this technology tends to methane-induced ageing. A loss of gas gain as well as a degradation of the energy resolution due to deposits on the electrodes was monitored. A qualitative and quantitative comparison between ageing in GEMs and proportional counters was performed.
The internet has often been considered a 'technology of freedom' – a nearly revolutionary tool believed to flatten social hierarchies and democratize access to media by 'giving voice' to everybody equally. Contradictory to this point of view, research has shown the existence of a 'digital divide,' the phenomenon that access to and use of the internet, as well as the outcomes derived from this use, correlate with pre-existing inequalities.
Based on ethnographic fieldwork among activists in Dakar, Senegal, this thesis analyzes how inequalities shape and are shaped by the relationships between activists and smartphones. Do smartphones indeed flatten social hierarchies, or are inequalities rather reproduced – or even reinforced – through them?
Computational workflow optimization for magnetic fluctuation measurements of 3D nano-tetrapods
(2021)
The detailed understanding of micro–and nanoscale structures, in particular their magnetization dynamics, dominates contemporary solid–state physics studies. Most investigations already identified an abundance of phenomena in one–and two–dimensional nanostructures. The following thesis focuses on the magnetic fingerprint of three–dimensional CoFe nano–magnets, specifically the temporal development of their hysteresis loop. These nano–magnets were grown in a tetrahedral pattern on top of a highly susceptible home–build GaAs/AlGaAs micro–Hall sensor using focused electron beam induced deposition (FEBID).
During the measurements, utmost efforts were employed to exemplify current best research practices. The data life cycle of the present thesis is based upon open–source data science tools and packages. Data acquisition and analysis required self–written automated algorithms to handle the extensive quantity of data. Existing instrumental-controlling software was improved, and new Python packages were devised to analyze and visualize the gathered data. The open–source Python data analysis framework (ana) was developed to facilitate computational reproducibility. This framework transparently analyses and visualizes the gathered data automatically using Continuous Analysis tools based on GitLab and Continuous Integration. This automatization uses bespoke scripts combined with virtualization tools like Docker to facilitate reproducible and device–independent results.
The hysteresis loops reveal distinct differences in subsequently measured loops with identical initial experimental parameters, originating from the nano–magnet’s magnetic noise. This noise amplifies in regions where switching processes occur. In such noise–prone regions, the time–dependent scrutinization reveals presumably thermally induced metastable magnetization states. The frequency–dependent power spectral density uncovers a characteristic 1/f² behavior at noise–prone regions with metastable magnetization states.
When performing transfer learning in Computer Vision, normally a pretrained model (source model) that is trained on a specific task and a large dataset like ImageNet is used. The learned representation of that source model is then used to perform a transfer to a target task. Performing transfer learning in this way had a great impact on Computer Vision, because it worked seamlessly, especially on tasks that are related to each other. Current research topics have investigated the relationship between different tasks and their impact on transfer learning by developing similarity methods. These similarity methods have in common, to do transfer learning without actually doing transfer learning in the first place but rather by predicting transfer learning rankings so that the best possible source model can be selected from a range of different source models. However, these methods have focused only on singlesource transfers and have not paid attention to multi-source transfers. Multi-source transfers promise even better results than single-source transfers as they combine information from multiple source tasks, all of which are useful to the target task. We fill this gap and propose a many-to-one task similarity method called MOTS that predicts both, single-source transfers and multi-source transfers to a specific target task. We do that by using linear regression and the source representations of the source models to predict the target representation. We show that we achieve at least results on par with related state-of-the-art methods when only focusing on singlesource transfers using the Pascal VOC and Taskonomy benchmark. We show that we even outperform all of them when using single and multi-source transfers together (0.9 vs. 0.8) on the Taskonomy benchmark. We additionally investigate the performance of MOTS in conjunction with a multi-task learning architecture. The task-decoder heads of a multi-task learning architecture are used in different variations to do multi-source transfers since it promises efficiency over multiple singletask architectures and incurs less computational cost. Results show that our proposed method accurately predicts transfer learning rankings on the NYUD dataset and even shows the best transfer learning results always being achieved when using more than one source task. Additionally, it is further examined that even just using one task-decoder head from the multi-task learning architecture promises better transfer learning results, than using a single-task architecture for the same task, which is due to the shared information from different tasks in the multi-task learning architecture in previous layers. Since the MOTS rankings for selecting the MTI-Net task-decoder head with the highest transfer learning performance were very accurate for the NYUD but not satisfying for the Pascal VOC dataset, further experiments need to varify the generalizability of MOTS rankings for the selection of the optimal task-decoder head from a multi-task architecture.