News

Vortrag von: Philippe Crackau
Titel: Examining Image Classifier Decisions Through the Lense of Probabilistic Relational Explanations – Extending the CoReX Framework by Probabilistic Logic Programming
Zeit: 11.06.2025 - 13:00 - 14:00
Ort: WE5/05.013 und Teams
Abstract:

Explainability and human interpretability are relevant topics of research due to
the ubiquity of AI systems in various fields through the rise of Deep Neural Net-
works and the growing awareness of their shortcomings. The CoReX-framework
is one such system, which uses Prolog to learn human readable First-order logic
rules from features informing the NNs decision-making.

In this thesis, the current Prolog-explainer Aleph will be replaced with a prob-
abilistic inductive logic programming extension in Prolog by Riguzzi et al. in
order to gain an additional layer of information by weighting the certainty of
the classification.

________________________________________________________________________________

Microsoft Teams Benötigen Sie Hilfe?

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 383 659 694 423 3

Kennung: Mv68nY9K

Vortrag von: Christoph Ubald Stößel 
Titel: Ein multimodales interaktives Lehr-Lern-System zur Vermittlung von KI-Kompetenzen für KMUs
Zeit: 31.07.2025 - 09:00 - 10:00
Ort: WE5/05.013 und Teams
Abstract:

Small and medium-sized enterprises (SMEs) represent a central business segment and play a crucial role for the German economy. Nevertheless, the increasingly important topic of artificial intelligence (AI) remains underrepresented by most SMEs, primarily due to a lack of necessary skills. This can be attributed to various inhibiting factors, one of which is the lack of access to information materials addressing the needs of the target group. As part of this study, a multimodal interactive tutor system was developed, to inform SME employees about AI and its practical applications, while also aiming to build relevant competencies. The content is tailored to the target group based on learning theories from adult education and multimedia learning. Furthermore, special emphasis is placed on ensuring a high level of usability. The development followed a design-based research approach, encompassing the phases of requirements analysis, implementation and evaluation. The system’s usability was evaluated using the System Usability Scale (SUS). Additional insights into content design were gained through qualitative supplementary questions. The results indicate that the system positively influences user perception, thanks to its wide range of interactive features and clear visual design. Moreover, the evaluation results offer valuable insights for potential improvements in a subsequent iteration of the system.

Information to participate via Teams:

________________________________________________________________________________

Microsoft Teams Benötigen Sie Hilfe?

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 334 724 880 415 0

Kennung: qv97Uf75 

Vortrag von: Amir Ebrahimi Moghaddam 
Titel: Revising CNN Models via Augmented Near Hits and Near Misses in Binary Image Classification
Zeit: 30.07.2025 - 16:00 - 17:00
Ort: WE5/05.013 und Teams
Abstract:

In this thesis, we propose a method for revising a baseline CNN model by augmenting images selected through Near Hits and Near Misses (NHMs) found in its misclassified predictions. Unlike traditional random augmentation, our approach aims to target the decision boundary and support more meaningful  model improvements. We experiment with different strategies for selecting misclassified images: randomly or based on low-confidence predictions close to the decision boundary. These misclassified  samples are then used to find NHMs( visually or semantically similar examples) which are augmented and added back into training. We compare the effects of NHM-based augmentation with random selection, measuring improvements in accuracy and generalization. Key challenges include ensuring diversity in augmentation, choosing the right model capacity, and determining how misclassified sample selection affects NHM quality. Our results aim to reveal whether model uncertainty and semantic similarity can guide more effective and interpretable model revisions.

Information to participate via Teams:

________________________________________________________________________________

Microsoft Teams Benötigen Sie Hilfe?

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 379 093 638 480

Kennung: jS3Mf7xf

Vortrag von: Kemal Tanirak
Titel: "Leben in Deutschland" als Serious Game: Vermittlung von deklarativem Wissen durch ein dialogbasiertes System auf Basis eines LLMs (MA CitH)
Zeit: 09.09.2025 - 14:00 - 15:00
Ort: WE5/05.005 und Teams
Abstract folgt

Information to participate via Teams:

________________________________________________________________________________

Microsoft Teams Benötigen Sie Hilfe?

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 344 663 060 513 1

Kennung: Yo7Qa7MW

Vortrag von: Max Brombach
Titel: Automated Test Case Generation for Bug Detection in Python Programs Using Large Language Models: A Comparative Study of End-to-End and Systematic Test Case Generation
Zeit: 27.08.2025 - 12:00 - 13:00
Ort: WE5/05.005 und Teams
Abstract folgt

Information to participate via Teams:

________________________________________________________________________________

Microsoft Teams Benötigen Sie Hilfe?

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 340 579 583 586 4

Kennung: 8aQ3Fp9V

Vortrag: Yuliia Kaidashova
Titel: Toward Short and Robust Contrastive Explanations for Image Classification by Leveraging Instance Similarity and Concept Relevance

Vortrag: Luisa Schneider
Titel: Re-Evaluating the Robustness and Interpretability of the Contrastive Explanation Method for Image Classification

Zeit: 10.09.2025 - 16:00 - 17:00
Ort: WE5/05.005 und Teams

________________________________________________________________________________

Microsoft Teams Benötigen Sie Hilfe?

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 379 093 638 480

Kennung: jS3Mf7xf

Vortrag von: Anna Thaler
Titel: Deep Spatiotemporal Learning for Facial Expression Classification: Using an Action Unit-Enhanced CNN-LSTM Network for Video-Based Classification of Pain and Disgust
Zeit: 25.09.2025 - 18:00 - 19:00
Ort: WE5/05.005 und Teams
Abstract:

Automated pain assessment from facial expressions faces a compounded challenge: pain must be inferred without self-report, facial actions overlap with those of disgust, and clinically realistic datasets with spontaneous expressions are typically small, which constrains model development and validation. This thesis examines whether commonly used deep learning methods with strong regularization can reliably distinguish the two aversive affective states of pain from disgust under conditions of limited data. It further investigates whether frame-level semantic cues derived from facial action units (AUs) enhance this discrimination. A CNN–LSTM architecture is implemented that applies transfer learning, using FaceNet (pre-trained on VGGFace2) as a backbone and fine-tuned via gradual unfreezing and discriminative learning rates on a subset of the proprietary PainFaceReader video dataset. An additional model variant incorporates concatenated frame-level AUs as auxiliary features. Evaluation through leave-one-participant-out (LOPO) cross-validation followed by statistical testing shows that both the baseline model and the AU-enhanced models distinguish pain from disgust significantly above chance (binomial test: p=0.002, and p=0.005 respectively) achieving accuracies, recall, and F1-scores of 56 % (baseline) and 57 % (AU-enhanced). However, the inclusion of AUs did not yield significant performance gains, and strong overfitting is still observed. The findings highlight that spontaneous facial expressions pose particular challenges for deep learning models, which tend to overfit when trained on small, clinically realistic datasets, thereby emphasizing the need for more data-efficient and adaptive methods.

________________________________________________________________________________

Microsoft Teams Benötigen Sie Hilfe?

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 367 104 110 913 9

Passcode: Hs3Lt7us

Vortrag von: Vu Thuy Giang Do
Titel: Latent Distribution Based Anomaly Detection for Multivariate Time-Series Data
Zeit: 26.09.2025 - 10:00 - 11:00
Ort: WE5/05.005 (nicht hybrid)
Abstract:

Anomaly detection in telemetry data plays a crucial role, as any abnormal data can compromise the space monitoring system. To date, a variety of architectural models exist that can detect anomalies by reconstructing normal and abnormal data under autoencoder architectures and determine the presence of anomalies based on reconstruction error; the greater the reconstruction error, the more likely it is that an anomaly is present. These models focus on reducing reconstruction error while learning high-dimensional time-series data, such as telemetry data. However, recent models based on reconstruction error to detect abnormal data may be neither sufficient nor high-performance for multivariate time-series data, as outlier data points can also be reconstructed too well, resulting in an unexpectedly low reconstruction error. Moreover, how the data is represented affects the performance of a machine learning approach. This thesis applies the LatentOut unsupervised algorithm designed by Angiulli et al. (2022) to take advantage of the latent space data in combination with the reconstruction error representation to detect outliers in such complex data. The final result shows that the anomalies lie in a tendency in the sparsest region of the combined space.