Contributions in Merlin

null

null

Scientific datasets often have hierarchical structure: for example, in surveys, individual participants (samples) might be grouped at a higher level (units) such as their geographical region. In these settings, the interest is often in exploring the structure on the unit level rather than on the sample level. Units can be compared based on the distance between their means, however this ignores the within-unit distribution of samples. Here we develop an approach for exploratory analysis of hierarchical datasets using the Wasserstein distance metric that takes into account the shapes of within-unit distributions. We use t-SNE to construct 2D embeddings of the units, based on the matrix of pairwise Wasserstein distances between them. The distance matrix can be efficiently computed by approximating each unit with a Gaussian distribution, but we also provide a scalable method to compute exact Wasserstein distances. We use synthetic data to demonstrate the effectiveness of our Wasserstein t-SNE, and apply it to data from the 2017 German parliamentary election, considering polling stations as samples and voting districts as units. The resulting embedding uncovers meaningful structure in the data.

Der Trend zum nachhaltigen Investieren ist ungebrochen. Doch die Kritik wächst. Die Eindämmung von Greenwashing und die politische Vereinnahmung von ESG drängen in den Vordergrund.

We incorporate externalities into the stable matching theory of two-sided markets. Extending the classical substitutes condition to markets with externalities, we establish that stable matchings exist when agent choices satisfy substitutability. We show that substitutability is a necessary condition for the existence of a stable matching in a maximal-domain sense and provide a characterization of substitutable choice functions. In addition, we extend the standard insights of matching theory, like the existence of side-optimal stable matchings and the deferred acceptance algorithm, to settings with externalities even though the standard fixed-point techniques do not apply.

null

Ethereum in its current version reaches a maximum transaction throughput of 15 transactions a second and thus suffers from a scalability problem. The Polygon proof-of-stake blockchain presents itself as an already active solution to this problem. Previous research focuses on the fairness in other proof-of-stake blockchains and the scalability issue in general. Our contribution is to provide a careful investigation of incentives and decentralisation in Polygon PoS. To this end, we analyse the scalability potential by having a look at transactions, usage and distribution of rewards to participants in the network. Our results indicate that Polygon PoS, as a cheap solution, can enhance the transaction throughput. Furthermore, the blockchain has a fairly good user adoption paired with climate-friendliness. However, in order to be the ultimate scaling solution, there is the need to double-down on incentives and increase performance by a lot. It can also be shown that certain participants get disproportionately more rewards than others, as seen by applying measures like the Gini and Nakamoto index to the data. Centralisation seems to be a problem throughout the network. In other words, we find that Polygon PoS at the current stage is lacking incentives and decentralisation and only early adopters of the blockchain can profit from it.

AI has the potential to revolutionize industries and improve daily life through the development of advanced machine learning (ML) and deep learning (DL) models. These models, such as chatbots and language models, use algorithms or artificial neural networks to recognize patterns and make decisions. ML involves training algorithms on large datasets to identify patterns and make decisions, while DL uses artificial neural networks composed of interconnected nodes called artificial neurons to process and transmit information. Neural networks can learn and make decisions by adjusting the connections between neurons based on input data. There are two types of ML and DL: unsupervised and supervised. Unsupervised learning involves using algorithms or neural networks to learn from data without labeled outcomes, while supervised learning involves training algorithms or neural networks on labeled data to make predictions or decisions. As AI becomes more advanced and widespread, it is important to have confidence in the decisions and actions of these systems. Trusted AI refers to the reliability and ethical behavior of AI systems. It is crucial to have a framework for evaluating the trustworthiness of different AI models to ensure their safe and responsible deployment. A taxonomy of pillars and metrics can be used to quantify the trustworthiness of AI models, allowing for a structured and comprehensive evaluation of their strengths and limitations. The following bachelor thesis aims to survey existing platforms, define requirements and develop a web app that allows the computation of the trustscore, pillarscores, metricscores of supervised and unsupervised and DL platform is extended to allow for user management, and the return of the trustworthiness levels via API endpoints.

null

Vision Transformers (ViTs) are powerful deep learning models and have recently made impressive strides in the computer vision field. However, vision transformers are not data efficient, and their high computational cost, quadratic in the number of tokens, currently limits their adoption in power- and computation-constrained applications. To improve the data and inference efficiency of ViTs, we explore two different paths. First, we notice that the tokens in ViTs do not take any inductive bias. We extract more fine-grained tokens (dubbed subtokens) from each token by expanding its channel dimension to spatial dimensions, and introduce convolutions or shifting on the subtokens to insert intra-token spatial priors. The subtoken convolution improves the classification accuracy for ViTs training from scratch by 2.21% on small datasets (Cifar100) and 1.14% on larger datasets (ImageNet-1K), and also shows faster convergence speed. Secondly, recent studies have shown that not all tokens are helpful for the final task, and ViTs can be made more efficient by pruning redundant tokens. However, active research is mostly focusing on high-level tasks like image classification. To extend the token pruning methods to more complex downstream tasks, we revisit the designs of token pruning and find three key components that lead to better performance: (1) the token selection should not be based on the class token, (2) a dynamic pruning rate is better than a static pruning rate, (3) preserving the feature map of all tokens is better than dropping tokens for all later layers. To this end, we propose SViT, a simple yet effective dynamic token selection scheme that selects and processes highly informative tokens while preserving a structured feature map, thus maintaining compatibility with downstream tasks. On the image classification task (ImageNet-1K), we improve the throughput of DeiT-S by 49% with only 0.4% accuracy drop. On object detection and instance segmentation tasks(COCO), we improve the inference speed by 32.5% with -0.3 box AP and no drop in mask AP.