[WS 22/23] Quality Assurance for Machine Learning

Please register for the course in MOSES or via email to secure a spot. Registration in ISIS will not be considered binding.


Learning outcomes:

Knowledge: Students will be familiar with the requirements imposed on AI systems by current and prospective regulatory frameworks. They will know different quality dimensions of AI systems and will have elaborated procedures and metrics for measuring quality along these dimensions and technical means for assuring high quality.

Skills: Students will acquire practical skills to debug data science applications by analyzing data and/or code. In addition, students will also acquire or refine skills to independently review and systematically structure the literature of a well circumscribed field in order to address a given set of questions, and will gain experience in presenting the outcome to a critical audience as well as in participating in scientific discussions.

Competencies: Students will be able to analyze the the risks and failure modes of a given AI system or product. They will be able to suggest procedures and tests to benchmark whether a given system/product will function as intended in a variety of settings and suggest ways to overcome unwanted system behavior. Students will also be able to discuss ethical, economic and other implications of failure modes of AI systems.


Content:

- motivating examples and use cases from the medical, automotive and finance domains
- taxonomy of use cases, associated risks and failure modes
- legal approaches to regulate AI systems for critical applications
- ethical considerations
- quality dimensions for AI/ML systems 
 * data quality 
 * model performance, generalization 
 * model robustness 
 * model fairness 
 * transparency   
 * uncertainty calibration 
 * model interpretation and types of explainability 
 * data privacy
- quantitative metrics and test to measure quality
- practical approaches to ensure/improve quality
- simulations and benchmarks
- current standardization efforts



Teaching methods:

Students will prepare a presentation to a specific topic based on a provided collection of published material. Student presentations will be framed by short lecture segments introducing, contextualizing and connecting the presented topics. Each course slot will contain discussion periods, in which active participation is fostered. Students will also conduct a group work to analyze a given AI system, and present their results to the seminar audience. Moreover, students will complete a homework.

Prerequisites: A BSc degree in Computer Science is recommended. Successful completion of an introductory module on ML such as "Machine Learning I" or "Machine Intelligence I" is recommended. Programming skills in at least one language (e.g., Matlab, Python, R) are required.

The module grade is calculated based on
1. The quality of a paper presentation (50%).
2. The quality of the presentation of a group work (25%).
3. The quality of a homework (25%).