Project group "XAI" wins the Alumni Informatik prize worth 1000 euros
The XAI project group has been working on the explainability of artificial intelligence (AI), and more specifically neural networks. These models often act as a "black box", which makes it difficult to trust their decisions. The concept of Explainable AI (XAI) aims to make these decisions more transparent. This not only strengthens trust in the models, but also helps to improve them. A classic example of the need for explainability is the use of X-ray images in medical diagnosis.
The project group has developed a framework in which new images can be uploaded and analyzed. The decision of the underlying model is examined and explained both visually and linguistically by an Large Language Model (LLM).
In image classification in particular, methods such as LRP, Lime and CAM are used to identify relevant image areas for the decision. After extensive evaluation by the project group, these methods proved to be the most reliable.
As an example, the popular game GeoGuessr was selected, in which the aim is to recognize random places on earth using Google Streetmaps images. This example is ideal for comparing the models' decisions with those of humans, as people usually use certain image content, such as license plates, landmarks or infrastructure, to identify countries. The LLM is used here to describe the differences and similarities found between model and human.