Overview
| Module | Master-Seminar / Advanced Seminar Course (MSc), |
| Prerequisites | You have to have good knowledge of logic (e.g. by passing Techniques in Artificial Intelligence (IN2062) or Logic (IN2049) or Model Checking (IN2050)or Efficient Algorithms II (IN2004)) AND of machine learning (e.g. by passing Machine Learning (IN2064) or Statistical Modeling and Machine Learning (IN2332)) |
| Language | English |
| Organisation | Mohammad Abdulaziz, Prof. Tobias Nipkow Contact Mohammad Abdulaziz for all course matters, or your advisor (once you have been assigned one). |
Content
The capabilities of systems based on machine learning has seen a boom, with the development of techniques like deep learning and the continuing reduction in the cost of computation. Due to these capabilities, these systems have proliferated in many applications and now have direct consequences on our day-to-day lives. However, one hurdle to the adoption of machine learning systems in safety-critical applications is that, until now, a lot of the practically relevant machine learning algorithms are poorly understood and, accordingly, they offer no provable safety guarantees on the learnt systems. Recent research has been remedying that situation by applying formal verification techniques to specific neural networks to verify that they do not violate certain safety properties, e.g. that a minor change in a picture would not cause a complete change in its classification by a neural network. In this seminar, each student will study and present a recent research paper in that area.
Required Work and Grading
Participants will each be assigned one topic, about which they will write a short paper (~10 pages) and give a presentation (~30–40 minutes). Additionally, each student must write a review of the papers of two other students. Furthermore, students are required to be present during other students’ presentations.
The final grade is composed as follows: 20 % reviews written + 40 % presentation + 40 % paper