A5005204077- Explainable Artificial Intelligence (XAI)
- Face and Expression Recognition
- Brain Tumor Detection and Classification
- Bayesian Modeling and Causal Inference
- Machine Learning and Data Classification
- Image Retrieval and Classification Techniques
- Fuzzy Logic and Control Systems
- Neural Networks and Applications
- Machine Learning in Healthcare
- Reinforcement Learning in Robotics
- Visual Attention and Saliency Detection
- Evolutionary Algorithms and Applications
- Recommender Systems and Techniques
- Advanced Image Fusion Techniques
- Bone and Joint Diseases
- Machine Learning in Materials Science
- Spondyloarthritis Studies and Treatments
- Image and Video Quality Assessment
- Hematological disorders and diagnostics
- Semantic Web and Ontologies
- Numerical Methods and Algorithms
- Web Data Mining and Analysis
Ghent University
2022-2024
Vlaams Instituut voor Biotechnologie
2023
Ghent University Hospital
2018-2022
University College Ghent
2022
Abstract Feature attribution maps are a popular approach to highlight the most important pixels in an image for given prediction of model. Despite recent growth popularity and available methods, objective evaluation such remains open problem. Building on previous work this domain, we investigate existing quality metrics propose new variants maps. We confirm finding that different seem measure underlying properties maps, extend larger selection metrics, datasets. also find metric results one...
We aimed to develop and validate a fully automated machine learning (ML) algorithm that predicts bone marrow edema (BME) on quadrant level in sacroiliac (SI) joint magnetic resonance imaging (MRI).A computer vision workflow automatically locates the SI joints, segments regions of interest (ilium sacrum), performs objective extraction, presence BME, suggestive inflammatory lesions, semicoronal slices T1/T2-weighted MRI scans. Ground truth was determined by consensus among human readers. The...
For smart decision making, user agents need live and historic access to open data from sensors installed in the public domain. In contrast a closed environment, for Open Data federated query processing algorithms, publisher cannot anticipate advance on specific questions, nor can it deal with bad cost-efficiency of server interface when consumers increase. When publishing observations sensors, different fragmentation strategies be thought depending how needs queried. Furthermore, both...
Feature attribution maps are a popular approach to highlight the most important pixels in an image for given prediction of model. Despite recent growth popularity and available methods, little attention is objective evaluation such maps. Building on previous work this domain, we investigate existing metrics propose new variants We confirm finding that different seem measure underlying concepts maps, extend larger selection metrics. also find metric results one dataset do not necessarily...
The black box problem in machine learning has led to the introduction of an ever-increasing set explanation methods for complex models. These explanations have different properties, which turn method selection: is most suitable a given use case? In this work, we propose unifying framework attribution-based methods, provides step towards rigorous study similarities and differences explanations. We first introduce removal-based attribution (RBAMs), show that extensively broad selection...
Because of their strong theoretical properties, Shapley values have become very popular as a way to explain predictions made by black box models. Unfortuately, most existing techniques compute are computationally expensive. We propose PDD-SHAP, an algorithm that uses ANOVA-based functional decomposition model approximate the black-box being explained. This allows us calculate orders magnitude faster than methods for large datasets, significantly reducing amortized cost computing when many need be
Deep Reinforcement Learning uses a deep neural network to encode policy, which achieves very good performance in wide range of applications but is widely regarded as black box model. A more interpretable alternative networks given by neuro-fuzzy controllers. Unfortunately, controllers often need large number rules solve relatively simple tasks, making them difficult interpret. In this work, we present an algorithm distill the policy from Q-network into compact controller. This allows us...
Deep Reinforcement Learning uses a deep neural network to encode policy, which achieves very good performance in wide range of applications but is widely regarded as black box model. A more interpretable alternative networks given by neuro-fuzzy controllers. Unfortunately, controllers often need large number rules solve relatively simple tasks, making them difficult interpret. In this work, we present an algorithm distill the policy from Q-network into compact controller. This allows us...