Responsible AI: Algorithmic fairness and explainability

• explain what AI means to discuss and identify the common assumptions we make, and the ones we should avoid when building AI responsibly,
• investigate, identify, and discuss issues with AI systems responsibly within and outside the community,
• give an overview of challenges and state-of-the-art in fairness and bias when building and deploying AI,
• diagnose bias in predictive and generative AI,
• be familiar with common sources of bias, and mitigate bias in predictive and generative AI by either adapting the modeling or employing mitigation strategies from algorithmic fairness,
• understand most common types of explainable AI models, and discuss their strengths and weaknesses,
• apply explainable AI models in real scenarios for predictive and generative AI, and interpret their explanations,
• reading research papers with both technical and ethical content, understanding them and presenting their content,
• presenting analysis and results in writing.

The course will consist of three parts:
The first part of the course will be about the epistemology of machine learning, model-fitting vs. artificial intelligence, Bayesian problems, and Generative AI. Next, the fairness component will review classical algorithmic fairness methods and their limitations and potential solutions. Finally, we will learn about different paradigms of explainable AI such as saliency and prototype based methods, their use cases, as well as their validation. We will discuss and experiment the possibility of explaining generative AI. As part of this, we will study the philosophical assumptions in explainable AI and review AI ethics.

02450/02456, We expect that students are able to implement and train machine learning models from custom data, including deep learning models.

The course will consist of lectures, discussions, and practical exercises. It will be structured around three topics, all starting out with an introduction of established knowledge, moving to reading and discussing state-of-the-art papers, and finally a practical case implementation building on the visited methods.