KogSys: Colloquium Topics in SS 2024 | VC Universität Bamberg

In this thread all special talks in our CogSys colloq will be announced.

Re: Collouium Topics in SS 2024

ó $a->name - $a->date

MA Verteidigung Daniel Gramelt: Explanatory interactive machine learning for welding seam quality assessment (MA AI, in cooperation with Porsche digital)
Zeit: 15.04.2024 - 12:00 - 13:00
Ort: WE5/05.013 und Teams

Information to participate via Teams:
Microsoft Teams Benötigen Sie Hilfe?
https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZDNhMWIxOGQtNGM1Zi00M2YwLThhMDctNmVkYzYyYWE5NDc2%40thread.v2/0?context=%7b%22Tid%22%3a%224f18ddfc-c31f-4597-afda-fa5a760bf3cf%22%2c%22Oid%22%3a%224d7684e9-7cb5-4b10-b8b3-cae1a82ac504%22%7d
Besprechungs-ID: 384 493 620 804
Kennung: qwfKMu

Re: Collouium Topics in SS 2024

ó $a->name - $a->date

Talk by David Cerna: Anti-uniﬁcation: Introduction, Applications, and Recent Results
Time: 11.04.2024 - 17:00-18:00
Where: WE5/05.013 and Teams

Anti-uniﬁcation is a method for symbolically generalizing formal expression. It was
introduced independently by Plotkin and Reynolds as an operation for inductive infer-
encing. Though conceptually simple, it is an eﬀective tool for abstraction and templat-
ing. Since the seminal work, the number of applications has grown tremendously with
uses in program analysis, program repair, library compression, automated reasoning, and
beyond. With the growth of applications, there has been an eﬀort to strengthen the the-
oretical foundations of the subject. In this talk, we introduce anti-uniﬁcation, overview
the existing applications, and discuss recent theoretical results concerning equational and
high-order anti-uniﬁcation.

TEAMS:
https://teams.microsoft.com/l/meetup-join/19%3ameeting_YWEwY2ZiNzYtOWIzMi00N2YwLTgwZTUtYmZiZTVlY2UzMDI4%40thread.v2/0?context=%7b%22Tid%22%3a%224f18ddfc-c31f-4597-afda-fa5a760bf3cf%22%2c%22Oid%22%3a%224d7684e9-7cb5-4b10-b8b3-cae1a82ac504%22%7d

Besprechungs-ID: 361 014 049 591
Kennung: 8TBetS

Presentation Slides:

DavidCerna_Presentation-April-2024.pdf

Re: Collouium Topics in SS 2024

ó $a->name - $a->date

Vortrag von:
Annabel Lindner (FAU), Doktorandin in der Informatik-Didaktik
Titel:
How to Deal with Transformative Topics in Computer Science Education: An Analysis Based on the Topic of Artificial Intelligence

Zeit: 19.06.2024 - 16:00 - 17:30
Ort: WE5/05.013 und Teams

Information to participate via Teams:

Join the meeting now

Meeting ID: 354 320 751 078

Passcode: CmdZ5p

Re: Collouium Topics in SS 2024

ó $a->name - $a->date

Vortrag von:
Lukas Gernlein
Titel:
Towards Explainable and Interactive Machine Learning for Multi-Label Classification

Abstract:

The opaque nature of neural networks hinders applicants from fully retracing the prediction done by deep learning classifiers. When employing neural networks in critical applications such as medical diagnosis or autonomous driving, predictions relying on incorrect decision boundaries can have fatal effects. Therefore, methods are needed to explain the prediction process of deep learning classifiers and to interact with these explanations to prevent malicious behaviour. Our work provides a method to interact with classifiers explanations and enable feedback through a novel loss function, producing robust and reliable models for complex multi-label tasks. We introduce the multi-label right for the right reason loss function (MuLRRR), which, besides right answers, ensures correct reasoning by using label-specific annotation masks on integrated gradient explanations corresponding to a classifier’s prediction, where outlying gradients are penalized. Our contribution is combining and extending two approaches in the XIML research. We use integrated gradients to explain a classifier’s prediction, which is constrained over the extended right for the right reason loss function. Additionally, we derive a metric to evaluate a model’s multi-label explanation performance. This multi-label explanation score (MuLX) uses integrated gradients for explanations and is a ratio of correctly highlighted parts of the input to incorrect activated locations, which indicates how well a model focuses on correct parts for its forecasts. The experiments showed that our MuLRRR loss function enhanced classification metrics and the proposed MuLX score compared to common binary cross entropy constraints.

Zeit: 29.05.2024 - 16:00 - 17:30
Ort: WE5/05.013 und Teams

Information to participate via Teams:

Join the meeting now

Meeting ID: 354 320 751 078

Passcode: CmdZ5p

Re: Collouium Topics in SS 2024

ó $a->name - $a->date

Vortrag von:
Leonhard Kestel, Masterarbeitszwischenpräsentation
Titel:

Tree Vitality Classification from Image Data - A Comparison of Multipectral Index Scores and Deep Learning Approaches

Zeit: 10.07.2024 - 16:00 - 17:30
Ort: WE5/05.013 und Teams

Information to participate via Teams:

Join the meeting now

Meeting ID: 354 320 751 078

Passcode: CmdZ5p

Re: Colloquium Topics in SS 2024

ó $a->name - $a->date

Dear all,

you are cordially invited to join the master's thesis defense of Maren Stümke: Identifying and Overcoming Student’s Misconceptions in Chemistry: An Intelligent Tutoring System to Teach Chemical Equations with Worked-out Analogical Examples

Date: 26.06.2024 - 09:00 - 10:00 s.t.
Location: WE5/05.013 (CogSys-Lab), non-hybrid

Abstract:

Intelligent Tutoring Systems (ITS) aim to provide students with comparable instructional advantages to that of a human tutor by responding to individual needs and providing immediate feedback. Originally dating back to the 1980s, they have proven useful in different academic and educational contexts, especially in STEM-related domains (VanLehn, 2011). However, an ITS helping students to understand the underlying system of building and solving chemical formulae is lacking although many students struggle with these concepts (Taskin and Bernholt, 2012). Consequently, this Master Thesis provides the implementation and some first testing of an ITS for chemical formulae, especially redox reactions. It follows the classic architecture of an ITS with a background knowledge comprising the rules of deriving equations, a student module identifying errors or misconceptions in students’ answers, a simple interface, and a feedback module where structurally analogous examples are selected. We chose analogies as a feedback method as they enable students to not only solve the example at hand but also grasp the underlying schemata. This is essential to truly understand chemical formulae. The ITS was tested in a qualitative evaluation study and proved to be processing reliably and being reasonably easy to understand.

Re: Colloquium Topics in SS 2024

ó $a->name - $a->date

Vortrag von: Lynn Greschner, Doktorandin am Lehrstuhl für Grundlagen der Sprachverarbeitung
Titel: The Interplay of Emotions and Convincingness in Argument Mining for NLP

Abstract: The field of argument mining as part of natural language processing (NLP) aims to model arguments in their overall argumentative structure, including evaluating their quality, for instance, by assessing if an argument is convincing. Emotions play an important role in the perception of convincingness. In my PhD project, we aim to understand the interplay of emotions and convincingness. In this presentation, I will give a brief introduction to NLP tasks. I present our approach to developing a methodology for annotating and modeling emotions and convincingness in arguments.

Zeit: DELAYED UNTIL BEGIN OF WS24/25
Ort: WE5/05.013 und Teams

Re: Colloquium Topics in SS 2024

ó $a->name - $a->date

Vortrag von: Christian Dormagen (MA Verteidigung)
Titel: Inductive Learning of System-level Process Behaviour Explanations - A Hybrid Machine Learning Approach for Process Mining

Abstract: Classical event logs in process mining often treat data as abstract tokens, which fail to capture the semantics of process behavior. In addition, the process of extracting event logs from relational databases introduces artifacts and information about the interactions between objects. To address these limitations, this work leverages Semantic Web technologies to construct a semantic event log that better captures the underlying process behavior and uses it to learn partial process models to explain patterns within the data. We first develop an OWL 2 DL-based process mining upper ontology based on the Object Centric Event Data metamodel, which provides a schema for representing process data. This ontology is instantiated with event data to create a semantic log, enriched with system-level events that encode specific patterns of interest across multiple traces, and DECLARE constraints to increase its expressiveness. Using DL-Learner, an inductive logic programming system for OWL ontologies, we then derive explanations for system-level patterns in the semantic log as partial DECLARE models. This approach allows us to validate our semantic log in a first

use case and to improve our understanding of system-level process behavior. By focusing on system-level behaviors and separating the data into those that are affected by them and those that are not, we avoid the problems that inductive logic programming for process mining usually faces in the need for negative examples. Our methodology is validated through application to synthetic and real event data, demonstrating the potential of integrating Semantic Web technologies with process mining to improve the accuracy and semantic expressivity of event logs. Initial results indicate that our approach can be used to represent process data and apply existing SemanticWeb technologies to learn partial DECLARE models, although further work is needed, particularly when it comes to reducing the complexity.

Zeit: 15.07.2024 - 14:00-15:00
Ort: WE5/05.013 und Teams

Information to participate via Teams:

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 391 029 280 150

Kennung: vBqYbQ

Re: Colloquium Topics in SS 2024

ó $a->name - $a->date

Vortrag von: Corbinian Weithmann (MA Zwischenpräsentation)
Titel: Erklärbare Fehlerklassifikation in der bildbasierten industriellen Qualitätskontrolle: Anwendung von CNNs und Identifikation von Klassenprototypen für die Entwicklung von Statoren.

Abstract:

Ausgangssituation: In der Abteilung EA-343 der BMW Group soll eine Künstliche Intelligenz (KI) zur Fehlerdetektion bei der Entwicklung von Elektromotor-Komponenten eingeführt werden. Als Pilotkomponente wird der Stator verwendet. Ziel ist es, die automatisierte Erkennung von Produktionsfehlern durch ein Convolutional Neural Network (CNN) zu testen.

Problem: Fehler wie Blasen oder Risse im Stator können während der Produktion oder in den Testzyklen auftreten und sind manuell schwer zu erkennen. Diese Fehler beeinträchtigen die Produktqualität erheblich. Ein CNN-Modell soll entwickelt werden, um diese Fehler zuverlässig zu identifizieren und zu klassifizieren, um die Produktqualität zu verbessern.

Methode: In der ersten Stufe wird mit Hilfe eines trainierten CNN-Modells entschieden, ob ein Fehler vorliegt. In der zweiten Stufe wird die Art des Fehlers bestimmt. Um die Qualität der Fehlererkennung mittels CNN zu verbessern, werden Methoden der erklärbaren KI (xAI) zur Analyse und Verbesserung des Modells eingesetzt.

Erwartete Ergebnisse und Nutzen: Die automatisierte Fehlererkennung durch CNN soll Ingenieuren in der Entwicklungsphase helfen, Fehler frühzeitig zu identifizieren und zu klassifizieren. Durch den Einsatz von xAI wird die Vertrauenswürdigkeit der Ergebnisse erhöht, sodass Ingenieure fundierte Entscheidungen zur Verbesserung von Design und Prozessen treffen können.

Zeit: 10.07.2024 - 15:00-16:00
Ort: WE5/05.013 und Teams

Information to participate via Teams:

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 360 811 900 514

Kennung: XYHGxq

Re: Colloquium Topics in SS 2024

ó $a->name - $a->date

Vortrag von: Barbara Blank (MA Verteidigung)
Titel: Umsetzung von NLP-Methoden für ein Intelligentes Tutorsystem zur Konzeptvermittlung mittels sokratischem Dialog – Erschließung des Themenfelds heimische Bäume für die Grundschule

Abstract:

Diese Arbeit beschreibt Schritt für Schritt, wie ein intelligentes tutorielles System in Form eines Chatbots mit der Domäne “Bäume des heimischen Waldes” unter Verwendung von Dialogflow erstellt und in eine Webpage integriert werden kann. Die Beschreibung erfolgt beispielhaft anhand von Einzelteilen des Chatbots. Die Gesamtheit aller verknüpften Einzelteile wird durch die Präsentation des vollständigen Entscheidungsbaumes veranschaulicht. Die Domäne des Chatbots umfasst den Themenbereich “Bäume des heimischen Waldes”. Da der Chatbot mit Menschen kommunizieren soll, ist er zur Verarbeitung menschlicher Sprache mittels NLP-Methoden fähig. Hierbei liegt der Fokus auf der Verarbeitung kindlicher Sprachmuster, da der Chatbot für den Einsatz in der Grundschule konzipiert ist. Der Konversationsstruktur des Chatbots liegt der sokratische Dialog als Methode des Unterrichtsgesprächs zu Grunde. Die Arbeit beinhaltet außerdem eine empirische Evaluation des Chatbots, der in der Grundschule bereits zum Einsatz gekommen ist.

Zeit: 22.07.2024 - 10:00-11:00
Ort: WE5/05.013 und Teams

Information to participate via Teams:

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 339 620 858 802

Kennung: 2mY5EU

Re: Colloquium Topics in SS 2024

ó $a->name - $a->date

Dear all,

From now on, I (Felix Haase) will be doing the management of the colloquium. If you have any questions about it, please reach me at felix.haase@uni-bamberg.de.

Please note that the default Teams Meeting details for the colloquium have changed due to organisational reasons.
These are the new details (also available in the Course overview):

Microsoft Teams Need Help?

Join the meeting now
Meeting ID: 323 348 325 661
Passcode: NJWRud

Re: Colloquium Topics in SS 2024

ó $a->name - $a->date

Vortrag von: Johannes Langer (MA Verteidigung)
Titel: A Neuro-symbolic Single-path Approach for Solving Raven’s Progressive Matrices

Abstract:

The ability to reason in abstract domains is a key component of human intelligence, and is tested in common intelligence tests such as Raven’s Progressive Matrices (RPM). In this work, I explore a Neuro-symbolic architecture for solving RPM problems with the goal of combining state of the art computer vision models with the ability to learn programs from only a few examples provided by Inductive Logic Programming. It aims to segment relevant parts of raw images, create disentangled representations of the image parts using a Variational Autoencoder, and learn the relational structure of the problem from these representations using ILP. The only labels required are the masks of relevant objects for the training of the instance segmentation module, which makes it easily transferable to adjacent domains. While a complete integration of the proposed system was not possible at this time, I am able to show promising results for each of the submodules.

Zeit: 26.08.2024 - 14:00-15:00
Ort: WE5/05.013 und Teams

Information to participate via Teams:

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 363 042 262 931

Kennung: ZwLiHL

Re: Colloquium Topics in SS 2024

ó $a->name - $a->date

Vortrag von: Satyam Pant (MA Verteidigung)
Titel: Model Revision for Images Based on Near-Misses and Near-Hits Explanation-Guided Augmentation

Abstract:

The growing reliance on artificial intelligence (AI) in critical real-world applications, such as healthcare diagnostics and industrial quality control, demands models with high predictive accuracy and reliability. However, developing such models is often hindered by the limited availability of high-quality labeled data and class imbalances, leading to challenges in achieving robust predictions, particularly for minority classes. To address these issues, this work introduces ExTRe (Explanation Guided Training Revision), a novel approach that enhances model performance through interactive, explanation-driven revision. ExTRe involves human experts correcting misclassified instances and examining near hits and near misses, thereby refining the model’s decision boundaries. The corrected labels are used to augment the training data, guiding targeted model improvements. The method was evaluated on two binary image classification tasks: a subset of MNIST (distinguishing between digits 3 and 8) and a real-world welding seams dataset (classified as good or bad). For the welding dataset, ExTRe improved the baseline accuracy from 73% (ROC AUC of 0.81) to 95% (ROC AUC of 0.92). Similarly, for MNIST, ExTRe increased the baseline accuracy from 93% (ROC AUC of 0.99) to 99% with a perfect ROC AUC of 1.00. Precision, recall, and F1-score also showed notable improvements. This study demonstrates that ExTRe provides a cost-effective method for refining model decision boundaries, particularly in data-scarce environments, while enhancing model interpretability. These results validate ExTRe’s effectiveness and lay the groundwork for future advancements in AI-driven model revision techniques.

Zeit: 16.09.2024 - 11:00-12:00
Ort: WE5/05.013 und Teams

Information to participate via Teams:

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 328 690 820 420

Kennung: WNJ4qe

Re: Colloquium Topics in SS 2024

ó $a->name - $a->date

Vortrag von: Leonhard Kestel (MA Verteidigung)
Titel: Tree-Vitality Classification from Image Data
A Comparison of Multispectral Indices and Deep Learning Approaches

Abstract:

With climate change as threat and staff shortage in forestry, tree detection and vitality classification from UAV-based data is becoming a crucial task in order to preserve and maintain our forests. However it is very challenging, particularly in deciduous and mixed forests. While most studies have relied on multispectral data (RGB + NIR + RE), recent advancements in deep learning models have shown promising results using only RGB data, which would make these kinds of analyses cheaper and more accessible. This study aims to compare the efficacy of these data sources as the foundation for tree crown detection and vitality classification. I trained models on 3-band RGB and 5-band multispectral data collected from forests around Bamberg, Germany. Two instance segmentation algorithms, Mask R-CNN and Mask2Former, were employed for Individual Tree Crown Delineation (ITCD) along with ResNet-50 for vitality classification. At an IoU (Intersection over Union) threshold of 0.5, the best segmentation algorithm (Mask R-CNN on RGB data) achieved a precision of 0.46, recall of 0.31 and thus an F1-score of 0.37 on the test set. The ResNet classifiers based on both types of data demonstrated an overall accuracy exceeding 0.8, with the RGB-trained model achieving the highest accuracy of 0.85. For vitality classification, the RGB model achieved an F1-score of 0.91 for the vital class, 0.79 for the dead class, and 0.48 for the degrading class. The multispectral model showed slightly lower performance with F1-scores of 0.89, 0.77, and 0.51 for the vital, dead, and degrading classes, respectively. Notably, multispectral data increased the recall for degrading trees by 13 percent points, which resulted in a slightly higher F1-score. Incorporating the genus label as additional information for the model significantly enhanced the precision of the dead class by 8% on RGB data and by 13% on multispectral data. This study demonstrates that RGB data is a viable alternative for tree health monitoring and that genus labels contain valuable information for vitality classification. Hence, a multi-label learning approach is proposed to leverage it. The aim of this thesis is to contribute to the overarching goal of an end-to-end application for the segmentation and classification of trees from airborne data, both RGB or multispectral.

Zeit: 26.09.2024 - 14:00-15:00
Ort: WE5/05.013 und Teams

Information to participate via Teams:

Jetzt an der Besprechung teilnehmen

Besprechungs-ID: 381 678 919 927

Kennung: 3yXC8S