Colloquium topics in SS 2023

Colloquium topics in SS 2023

de Jonas-Dario Troles -
Número de respuestas: 5

In this thread we will announce all topics of the upcoming colloquiums. Looking forward to see you for discussions sonrisa

En respuesta a Jonas-Dario Troles

Re: Colloquium topics in SS 2023

de Jonas-Dario Troles -
Tuesday 19.09.2023 10:00am:

Felix Hempel's Master's thesis defence: Explanatory and Interactive Machine Learning with Counterfactual Explanations for Ordinal Data

Abstract:
When powerful machine learning methods are applied to sensitive domains such as medicine, experts must be able to rely on them when making critical decisions. However, many state-of-the-art machine learning models lack interpretability and directability, two essential properties that support trust in a system. In recent years, much effort has been put into methods that make a given model’s decision understandable. In addition, human-in-the-loop approaches involve domain experts in the learning process, allowing them to actively steer the model toward their desires and goals.
CAIPI is an explanatory and interactive Machine Learning approach that incrementally improves a model by reacting to user feedback during training. While the feedback of other interactive machine learning methods is focused on labels, CAIPI also acquires feedback on explanations. This makes training more effective and encourages (or discourages) users to trust the model. CAIPI has been mainly tested on image data using Local Interpretable Model-agnostic Explanations (LIME). Although LIME is model-agnostic, its capabilities on tabular data are limited.
This masters thesis introduces ordinal CAIPI for explanatory interactive learning on ordinal data. OCAIPI uses counterfactual explanations to explain its predictions. Counterfactual explanations are hypothetical if-then scenarios that reveal causal relationships by showing what changes to the input lead to a different output. They not only reveal the important features that lead to a decision but also provide concrete suggestions for potential changes to steer the outcome in a desired direction. OCAIPI features a graphical user interface through which the user can interact with predictions and their explanations if necessary. The model reacts on the users feedback and is trained incrementally.
In experiments it could be demonstrated that OCAIPI effectively accelerates training, particularly when dealing with data drift. Additionally, OCAIPI exposes deceptive features within the data that might have otherwise gone unnoticed. Moreover, users can effectively shift the model’s focus away from these artifacts solely by employing
explanation corrections.
En respuesta a Jonas-Dario Troles

Re: Colloquium topics in SS 2023

de Jonas-Dario Troles -
Master's thesis defense by Richard Nieding on 9th of October 11 am

Title: Unsupervised Machine Learning via Feature Extraction and Clustering to Classify Tree Species from High-Resolution UAV-based RGB Image Data

Abstract:
Forest fires, prolonged drought periods, and bark beetle infestation increasingly stress tree health. Drones and technological advancements are reducing the workload of arborists and foresters by speeding up forested areas’ analysis and inventory management to gain information about these stressed areas. One aspect of this analysis examines tree species distribution in a forest area. Hence, this work proposes an innovative unsupervised tree species classification ensemble to support tree species assessment. The ensemble uses an individual tree crown delineation model based on a mask region-based convolutional neural network to obtain single tree crown images from an orthomosaic. These images are fed into an experiment that uses different preprocessing steps, convolutional neural networks for feature extraction, dimensionality reduction techniques, and clustering techniques in an automated manner. The optimal method combination with contrast limited adaptive histogram equalization (CLAHE), DenseNet, principal component analysis, and k-means++ achieves a weighted F1-score of 0.79 on the FORTRESS dataset, which is only a little worse compared to supervised approaches. The qualitative results reveal that the method combination with CLAHE, DenseNet, uniform manifold approximation and projection, and agglomerative clustering generally performs well. Also, the results show that the ensemble is transferable and less stringent in classifying tree species in previously unknown forest areas. However, the proposed ensemble needs improvements in classifying classes with small sample sizes and different deciduous tree species. This work highlights the promise and importance of developing and using an unsupervised approach to classify tree species.
En respuesta a Jonas-Dario Troles

Re: Colloquium topics in SS 2023

de Jonas-Dario Troles -
Today from 11:00-12:00 in WE5/05.013:
Julian Holger Fehr Master's Thesis
Titel: Explainable Link Prediction in Knowledge Graphs with Text Literals - Enhancing the MINERVA Algorithm with SBERT
Abstract: In Knowledge Graphs (KGs), information is stored and as triples composed of entities and relations. However, KGs are sometimes incomplete, which has led to the development of several approaches for Link Prediction - the task of finding unknown relations - that allow the inference of new knowledge from existing knowledge. Text Literals, which contain the names and descriptions of the entities and relations, are however not considered in most approaches. This paper presents an approach for Link Prediction that allows the use of Text Literals as part of the learning process while maintaining the explainability of the model. This is intended to incorporate previously unused contextual information into link prediction in order to improve the results. As a basis for this approach, MINERVA, an explainable Link Prediction approach based on Reinforcement Learning, is used. SBERT is used to generate the representation of the textual contents while InputXGradient contributes to maintain the explainability of the model. The approach is evaluated using the LiterallyWikidata data set.