Semantische Suche
Donnerstag, 16. Februar 2023, 10:00 Uhr
iCal (Download)
Webkonferenz: https://kit-lecture.zoom.us/j/67744231815
| Vortragende(r) | Christoph Batke |
|---|---|
| Titel | Improving Document Information Extraction with efficient Pre-Training |
| Vortragstyp | Proposal |
| Betreuer(in) | Edouard Fouché |
| Vortragsmodus | online |
| Kurzfassung | SAP Document Information Extraction (DOX) is a service to extract logical entities from scanned documents based on the well-known Transformer architecture. The entities comprise header information such as document date or sender name, and line items from tables on the document with fields such as line item quantity. The model currently needs to be trained on a huge number of labeled documents, which is impractical. Also, this hinders the deployment of the model at large scale, as it cannot easily adapt to new languages or document types. Recently, pretraining large language models with self-supervised learning techniques have shown good results as a preliminary step, and allow reducing the amount of labels required in follow-up steps. However, to generalize self-supervised learning to document understanding, we need to take into account different modalities: text, layout and image information of documents. How to do that efficiently and effectively is unclear yet. The goal of this thesis is to come up with a technique for self-supervised pretraining within SAP DOX. We will evaluate our method and design decisions against SAP data as well as public data sets. Besides the accuracy of the extracted entities, we will measure to what extent our method lets us lower label requirements. |
Freitag, 3. März 2023, 11:30 Uhr
iCal (Download)
Ort: Raum 348 (Gebäude 50.34)
Webkonferenz: https://sdqweb.ipd.kit.edu/wiki/SDQ-Oberseminar/Microsoft_Teams
| Vortragende(r) | Janek Speit |
|---|---|
| Titel | Automated Classification of Design Decision in Software Architecture Documentation |
| Vortragstyp | Masterarbeit |
| Betreuer(in) | Jan Keim |
| Vortragsmodus | in Präsenz |
| Kurzfassung | Die Softwarearchitekturdokumentation (SAD) ist ein integrales Artefakt eines Softwareprojektes. Um die Qualität von SADs zu verbessern und nachgelagerte Aufgaben zu unterstützen, ist eine automatische Klassifizierung dieser Entwurfsentscheidungen erstrebenswert. In dieser Arbeit implementieren und evaluieren wir einen Ansatz zur automatischen Identifikation und Klassifizierung von Entwurfsentscheidungen auf der Grundlage einer feingranularen Taxonomie, bei der wir eine hierarchische Klassifikationsstrategie mit dem Einsatz von Transfer-Lernen durch vortrainierter Sprachmodelle kombinieren. Der Beitrag dieser Arbeit besteht darin, den Vorteil einer hierarchischen Klassifikationsstrategie für die automatische Klassifikation von Entwurfsentscheidungen gegenüber einem nicht-hierarchischen Ansatz zu untersuchen. Außerdem untersuchen und vergleichen wir die Effektivität verschiedener vortrainierter Sprachmodelle. |
| Vortragende(r) | Stefanie Fischer |
|---|---|
| Titel | Faster Feedback Cycles via Integration Testing Strategies for Serverless Edge Computing |
| Vortragstyp | Masterarbeit |
| Betreuer(in) | Robert Heinrich |
| Vortragsmodus | in Präsenz |
| Kurzfassung | Serverless computing allows software engineers to develop applications in the cloud without having to manage the infrastructure. The infrastructure is managed by the cloud provider. Therefore, software engineers treat the underlying infrastructure as a black box and focus on the business logic of the application. This lack of inside knowledge leads to an increased testing difficulty as applications tend to be dependent on the infrastructure and other applications running in the cloud environment. While isolated unit and functional testing is possible, integration testing is a challenge, as reliable results are often only achieved after deploying to the deployment environment because infrastructure specifics and other cloud services are only available in the actual cloud environment. This leads to a laborious development process. For this reason, this thesis deals with creating testing strategies for serverless edge computing to reduce feedback cycles and speed up development time. For evaluation, the developed testing strategies are applied to Lambda@Edge in AWS. |
Donnerstag, 9. März 2023, 10:00 Uhr
iCal (Download)
Webkonferenz: https://kit-lecture.zoom.us/j/67744231815
| Vortragende(r) | Dan Jia |
|---|---|
| Titel | Reinforcement Learning for Solving the Knight’s Tour Problem |
| Vortragstyp | Proposal |
| Betreuer(in) | Edouard Fouché |
| Vortragsmodus | online |
| Kurzfassung | The knight’s tour problem is an instance of the Hamiltonian path problem that is a typical NP-hard problem. A knight makes L-shape moves on a chessboard and tries to visit all the squares exactly once. The tour is closed if a knight can finish a complete tour and end on a square that is a neighbourhood of its starting square; Otherwise, it is open. Many algorithms and heuristics have been proposed to solve this problem. The most well-known one is warnsdorff’s heuristic. Warnsdorff’s idea is to move to the square with the fewest possible moves in a greedy fashion. Although this heuristic is fast, it does not always return a closed tour. Also, it only works on boards of certain dimensions. Due to its greedy behaviour, it can get stuck into a local optimum easily. That is similar to the other existing approaches. Our goal in this thesis is to come up with a new strategy based on reinforcement learning. Ideally, it should be able to find a closed tour on chessboards of any size. We will consider several approaches: value-based methods, policy optimization and actor-critic methods. Compared to previous work, our approach is non-deterministic and sees the problem as a single-player game with a tradeoff between exploration and exploitation. We will evaluate the effectiveness and efficiency of the existing methods and new heuristics. |
Freitag, 17. März 2023, 11:30 Uhr
iCal (Download)
Ort: Raum 348 (Gebäude 50.34)
| Vortragende(r) | Rakan Al Masri |
|---|---|
| Titel | Generating Causal Domain Knowledge for Cloud Systems Monitoring |
| Vortragstyp | Bachelorarbeit |
| Betreuer(in) | Pawel Bielski |
| Vortragsmodus | in Präsenz |
| Kurzfassung | While standard machine learning approaches rely solely on data to learn relevant patterns, in certain fields, this may not be sufficient. Researchers in the Healthcare domain, have successfully applied causal domain knowledge to improve prediction quality of machine learning models, especially for rare diseases. The causal domain knowledge informs the machine learning model about similar diseases, thus improving the quality of the predictions.
However, some domains, such as Cloud Systems Monitoring, lack readily available causal domain knowledge, and thus the knowledge must be approximated. Therefore, it is important to have a systematic investigation of the processes and design decision that affect the knowledge generation process. In this study, we showed how causal discovery algorithms can be employed to generate causal domain knowledge from raw textual logs in the Cloud Systems Monitoring domain. We also investigated the impact of various design choices on the domain knowledge generation process through systematic testing across multiple datasets and shared the insights we gained. To our knowledge, this is the first time such an investigation has been conducted. |
Freitag, 24. März 2023, 11:30 Uhr
iCal (Download)
Ort: Raum 348 (Gebäude 50.34)
| Vortragende(r) | Yannick Ettwein |
|---|---|
| Titel | Explainable Artificial Intelligence for Decision Support |
| Vortragstyp | Bachelorarbeit |
| Betreuer(in) | Vadim Arzamasov |
| Vortragsmodus | in Präsenz |
| Kurzfassung | Policy makers face the difficult task to make far-reaching decisions that impact the life of the the entire population based on uncertain parameters that they have little to no control
over, such as environmental impacts. Often, they use scenarios in their decision making process. Scenarios provide a common and intuitive way to communicate and characterize different uncertain outcomes in many decision support applications, especially in broad public debates. However, they often fall short of their potential, particularly when applied for groups with diverse interests and worldviews, due to the difficulty of choosing a small number of scenarios to summarize the entire range of uncertain future outcomes. Scenario discovery addresses these problems by using statistical or data-mining algorithms to find easy-to-interpret, policy-relevant regions in the space of uncertain input parameters of computer simulation models. One of many approaches to scenario discovery is subgroup discovery, an approach from the domain of explainable Artificial Intelligence. In this thesis, we test and evaluate multiple different subgroup discovery methods for their applicabilty to scenario discovery applications. |
| Vortragende(r) | Georg Gntuni |
|---|---|
| Titel | Streaming MMD Change Detection |
| Vortragstyp | Bachelorarbeit |
| Betreuer(in) | Florian Kalinke |
| Vortragsmodus | in Präsenz |
| Kurzfassung | Kernel methods are among the most well-known approaches in data science. Their ability to represent probability distributions as elements in a reproducing kernel Hilbert space gives rise to maximum mean discrepancy (MMD). MMD quantifies the dissimilarity of two distributions and allows powerful two-sample tests on many domains. One important application of general two-sample tests is change detection in data streams: Here, one tests the null hypothesis that the distributions of data within the stream do not change versus the alternative hypothesis that the distributions do change; a change in distribution then indicates a change point. The broad applicability of kernel-based two-sample tests renders their use for change detection in data streams highly desirable. But, their quadratic runtime complexity prohibits their application. While approximations for kernel methods that reduce their runtime in the static setting exist, their application to data streams is challenging.
In this thesis, we propose a novel change detector, RADMAN, which leverages the random Fourier feature-based kernel approximation to efficiently detect changes in data streams with a polylogarithmic runtime complexity of O(log^2 n) per insert operation, with n the total number of observations. The proposed approach runs significantly faster than existing methods but obtains similar result quality. Our experiments on synthetic and real-world data sets show that it performs better than current state-of-the-art approaches. |