Interactive Explainable AI
This project provides a toolbox that enables domain experts to inspect and correct AI models through their local explanations, demonstrated for automated seizure detection.
The explainable AI project has successfully enhanced the reliability of complex AI models in practical settings by providing clear insight into the model's decision process. The implemented approach locally explains the decisions of trained AI models, enabling domain experts to review and validate the explanations. When incorrect explanations were identified, the model's performance and trustworthiness were improved through our developed correction mechanism (IXAI). The strategy proved effective in epilepsy detection, with neurologists validating the results. The developed toolbox is now accessible to domain experts, requiring minimal annotations to improve their models across different use cases. This implementation satisfies the human oversight requirements of the European AI Act for high-risk systems. The project delivered significant improvements, including implementations of five major explanation methods (Occlusion, SHAP, DeepLift, Integrated Gradients, Expected Gradients), and demonstrated a 5% accuracy increase on standard Imagenet datasets through explanation correction.
Contact
Contact: Maarten De Vos (KU Leuven)
PhD: Joran Michiels (KU Leuven)
Research groups: KU Leuven - Stadius