We need to understand and explain what our models have learned. Concept-based XAI is a step on the way to get there.

In this tutorial, researchers and practitioners of DNNs are invited to learn which methods from explainable artificial intelligence (XAI) are ready at hand to gain insights into computer vision and language models. Our technical deep dive will highlight emerging concept-based explainability techniques (C-XAI), followed by a practical session including a discussion of participant-contributed use-cases.

An Overview

With the advances of deep learning, in particular in computer vision and natural language processing, came the pressing need to make their inner workings more explainable: Not only do ethical requirements arise in critical areas but better insights can also help to steer research on how to further improve the models. The resulting emerging field of explainable AI (XAI) has since introduced a vast range of methods and tools that can assist with diverse explainability problems—possibly overwhelming novices and practitioners. This half-day tutorial aims to lift the curtain, and both provide an overview on XAI problems and techniques, as well as a practical guide to method selection, practiced in a hands-on exploration and discussion part. As a highlight, a special focus will be given to the emerging field of concept-based explanations of latent spaces, as well as future challenges that could be a starting point for entering the research field. We welcome both junior researchers as well as experts from areas using or developing DNNs.

Techniques and Methods Covered

The following explainability methods will be introduced:

• Ante-hoc explainability techniques: transparent and hybrid models.
• Post-hoc model distillation: Techniques to approximate (parts of) a trained DNN with an explainable model.
• Post-hoc local feature importance methods: These identify features in a given input that have a high attribution to the DNN output.
• Concept-based explanations, following our survey:
  • Ante-hoc explainable concept-bottleneck models: A class of hybrid models inserting interpretable intermediate outputs into the DNN architecture.
  • Post-hoc concept mining: Methods to find and explain prevalent latent space features of DNNs which can help to find out biases.
  • Post-hoc concept extraction: Localize user-provided natural language concepts within the DNN latent space.
• Mechanistic Interpretability: Probing, sparse autoencoders, causal abstractions.

This knowledge is deepened by a practical exercise part that aims to transfer the provided guidelines to real-world problems.

Organizational

The tutorial will take place as part of the European Conference on Artificial Intelligence 2025 (ECAI2025), Oct 25—30, 2025, Bologna, Italy.

• When: exact date tba (within Oct 25—30, 2025)
• Where: Bologna, Italy. Check out the ECAI2025 venues.
• Format: half-day tutorial at ECAI2025 both with lecture and practical parts.
• Target audience: We target both researchers and practitioners seeking to integrate explainability into their work.
• Prerequisites: Some basic knowledge about deep learning will be required; experience with pytorch will be helpful for the practical part, but not required.
  Both novices and experts are welcome to join.
• What to bring: Feel free to bring and share your explainability problem during the discussion part.