 Hi, I'm Rina from the Artificial Intelligence Group at TU Dortmund, Germany. Together with Jan Niko and Alex, I work on debugging tools to support trustworthy machine learning. As most of you know, trustworthy machine learning is especially important for safety-critical areas like medical applications. The main problem with machine learning errors is how they accumulate over time, making it hard to pinpoint a single root cause. At the first steps towards developing automated error detection tools last year, Jan and I conducted a small survey on errors in machine learning pipelines to determine which errors could be detected. We generated a web-based questionnaire where 85 people from academia and industry participated. We asked the participants at which stage in the data science project they observed the most serious errors. On the x-axis of the graph, we see different stages of a machine learning pipeline. Here, the data pre-processing appears to be one of the most critical stages in both academia and industry. With those results, we decided to first focus on error detection tools for the data pre-processing stage. Here different errors can occur, like wrong handling of missing values or outliers. Now the question is, how can we automatically detect errors in the data pre-processing stage? As an easy-to-understand example, let's have a look at the normalization error to get a rough idea of how we plan to detect errors. In general, we want to extract automated checks from expert knowledge. On the left-hand side, we have a code snippet. So first the data is loaded, then some pre-processing steps are applied to the data before a support vector machine is trained and tested. Every data science expert would tell us that an SVM classification needs normalized data. Otherwise, we get wrong results for the future importance. This might not be clear for beginners. Our debugging tool could now check different parts of the machine learning program to figure out if normalization was applied or not. First, the debugger could run a test of the input data set to figure out if the data was already normalized. We could also statically check if we find function calls like scaling used for normalization. Some machine learning algorithms support default normalization. This could also be checked via the list of parameters of the SVM function. So from one expert rules, we need to extract several checks to construct a useful debugger. As our goal is to develop error detection tools for machine learning pipelines, our next step is to gather more reproducible code examples to test our tools. That's why we appreciate your contribution. We will also share our code examples and tools in an open repository. If you have executable code examples, please contact me via the email below. Thanks a lot.