This Week I Learned - Week 41 2021


It was my pleasure to represent the UniTheater Karlsruhe at the Multiplier training of the Federal Association for Amateur Theatres. I did a three-day intensive course on stage fighting and choreography with the amazing Florian Federl. Essentially but seemingly just coming up is the topic of consent in amateur theater.


I blogged about the university of the future. This follows the arguments by (EduTechAsia2019, 2019):

  • Standardised curriculum does not cater to individual needs
  • Limited 1-to-1 tutor time available in higher education
  • Grading & assessment is time-consuming, with an over-reliance on written tests
  • Personalized communication is almost impossible due to scale
  • Selecting the best students from a large application pool
  • Increasing dropout rates
  • The need to effectively combat plagiarism and ensure authorship

In Germany, these efforts are currently culminating in funding AI in Higher Education with 133 Mio. €. by the Federal Ministery for Research.

I read Hendrycks et al. (2021) paper about AI safety and robustness, which I think is a very readable introduction. When talking about robustness, problems can be one of two categories * Black Swans, which happen very rarely and at random with no real outer influence * Adversarial Attacks, which are carefully crafted to pose a deceptive threat to model outcomes

Robustness means many things but I want to highlight just a couple of thoughts from the paper: * Models should assess their domain of competence accurately and return calibrated uncertainty measures, so models are not overconfident in their competence. * Systems are more and more trained on data scraped from online data. Attackers can poison this data by manipulating data and publishing it online. When models are trained on this data, that adversaries can poison * Alignment is a specification problem as human goals are very hard to specify and even harder to measure

This concludes in the Swiss Cheese Model of AI Safety where unperfect Protection of the Layers of External Safety (utilizing ML for the safety of ML), Monitoring, Robustness, and Alignment combined can increase the security and safety of ML models.

Swiss Cheese Model of AI Safety


Something in the line of learning in lecture does not work: Students prefer lecture with PowerPoint even though it has “no measurable influence on course performance and minimal effect on grades”(Hill et al., 2012) and short-term or long-term memory of lecture content. (Nouri and Shahid, 2005) A newer study even found that “providing slides to students impacts negatively on their academic performance.” Leon and GarciaMartinez (2021)


Dan Hendrycks, Nicholas Carlini, John Schulman, and Jacob Steinhardt. Unsolved problems in ml safety. arXiv preprint arXiv:2109.13916, 2021.

Andrea Hill, Tammi Arford, Amy Lubitow, and Leandra M Smollin. “i’m ambivalent about it” the dilemmas of powerpoint. Teaching Sociology, 40(3):242–256, 2012.

Samuel P León and Inmaculada García-Martínez. Impact of the provision of powerpoint slides on learning. Computers & Education, 173:104283, 2021.

Hossein Nouri and Abdus Shahid. The effect of powerpoint presentations on student learning and attitudes. Global Perspectives on Accounting Education, 2:53, 2005.

EduTech Asia 2019. How artificial intelligence is disrupting education. 2019. URL: