• Ingold Rolf

• Abou Khaled Omar
• Mugellini Elena

With the increasing use of automation, users tend to delegate more and more tasks to the machines. Complex systems are usually developed with an Artificial Intelligence (AI) and can embed different kinds of models and algorithms including Machine Learning and Deep Learning, which make these systems difficult to understand for the user. This assumption is for instance particularly true in the field of automated driving since the level of automation is increasing or in the health domain where more and more sophisticated AI powered diagnostic tools are used every day. In order to better understand how AI works and build trust in the decisions made by AIs, new techniques in the field are emerging that are referred to as Explainable AI (XAI) [1].
These techniques are intended to make AI transparent and the contents of the "black boxes" accessible. The main purposes of this transparency are to:
- understand the functioning of algorithms and AIs in order to optimize their design and architecture, their features but also to understand and interpret the results
- increase human confidence in systems
- increase and improve cooperation between agents
As shown by [2], providing appropriate explanations to the user increases the user’s confidence in the system and thus allows for better human-IA collaboration.

The goal of this seminar is to investigate the field of Explainable AI (XAI) with a particular focus on the perspective of human interaction since it has not been sufficiently studied in existing explainable approaches [1] [3]. The seminar will address the topics related to the design of human-computer interfaces for XAI. Effective knowledge transfer through an explanation depends on a combination of AI algorithms used, explanation dialogues, and interfaces that can accommodate explanations.
Questions like “what kind of explanation do we need”, “what an explanation should look like? ”, “which is the best trade-off between performance and explainability we want to achieve”, “how granular should the explanations be” and “how to evaluate explanations” will be investigated in this seminar.

This seminar will help the students to improve their research and practical abilities. It will have a strong practical component as students will investigate existing applications as well as develop new concepts in the aforementioned domains.

- Identify and illustrate existing approaches in Explainable AI.
- Discuss and compare different methods for increasing system interpretability and transparency.
- Evaluate and select the best existing interactions and interfaces for intelligibility
- Identify and describe different ways of evaluating system explanability, accountability and intelligibility
- Identify and describe how to design interfaces to increase AI system predictability

MSc-CS BENEFRI - (Code Ue: 33823 / Tracks: T3, Code Ue: 63823 / T6) The exact date and time of this course as well as the full course list can be found under http://diuf.unifr.ch/drupal/mcs/program/courses-timetable/courses.