After successful completion of the course, students are able to...

• understand and critically reflect various research papers
• to develop topics scientifically 
• present their findings in front of an audience

In this seminar, we will explore cutting-edge research topics at the intersection of computer science, machine learning and physics. Together we will discuss and analyze papers with the aim to build an understanding beyond the level of mere applications of some currently emerging technologies. For this seminar there is no strict syllabus; the following is a list of possible interesting topics.

The following topics are subject of this course:

• Theory of deep learning
  Despite the huge recent success of deep learning, the behavior of deep neural networks (DNNs) is surprisingly poorly understood, and in many ways DNNs remain a black box. As a consequence, the design of new network architectures tailored to specific applications remains an ad-hoc process, and interpretation and analysis of (incorrect) results proves challenging. Some of the approaches to tackle this problem are: describing DNNs from the point of view of dynamical systems and control theory [1], and analyses based on mean-field theory [2, 3]. The latter approach allows, for example, for statements about the expressiveness and trainability of certain network architectures, and reveals intriguing connections to statistical physics.

• Hybrid modelling techniques
  Although it seems that DNNs can be tailored to describe almost any system given enough training data, naive approaches typically neglect prior knowledge about the inner workings of the modeled system. Existing domain knowledge (e.g., physical laws describing the time evolution of a dynamical system) can, however, be used to constrain the admissible solution space, leading to a decrease in required training data and to increased predictive accuracy. In combination with classical numerical techniques DNNs can be used to find data-driven solutions of differential equations [4] or for data-driven discovery of the governing equations themselves [5].

• Quantum computing and (post-) quantum cryptography
  As the first fully functional (however small) quantum processors are starting to appear [6], researchers are discovering an increasing number of applications that are potentially susceptive to a quantum speed-up, including many machine-learning algorithms [7]. Another active research field triggered by the advent of quantum computing is finding so-called quantum-resistant cryptographic algorithms, which cannot be broken by Shor’s algorithm.

1. Liu, Guan-Horng, and Evangelos A. Theodorou. ArXiv:1908.10920 [Cs, Eess, Stat], September 28, 2019. http://arxiv.org/abs/1908.10920.
2. Raghu, Maithra, Ben Poole, Jon Kleinberg, Surya Ganguli, and Jascha Sohl- Dickstein. arXiv:1606.05336 [Cs, Stat], June 18, 2017. http://arxiv.org/abs/1606.05336.
3. Schoenholz, Samuel S., Justin Gilmer, Surya Ganguli, and Jascha Sohl-Dickstein.arXiv:1611.01232 [Cs, Stat], April 4, 2017. http://arxiv.org/abs/1611.01232.
4. Raissi, Maziar, Paris Perdikaris, and George Em Karniadakis. arXiv:1711.10566 [Cs, Math, Stat], November 28, 2017.http://arxiv.org/abs/1711.10566.Raissi, Maziar, Paris Perdikaris, and George Em Karniadakis. arXiv:1711.10561 [Cs, Math, Stat], November 17, 2017.http://arxiv.org/abs/1711.10561.
5. Arute, Frank, Kunal Arya, Ryan Babbush, et al. Nature 574, no. 7779 (October 2019): 505–10. https://doi.org/10.1038/s41586-019-1666-5.
6. Schuld, M., I. Sinayskiy, and F. Petruccione. arXiv:1409.3097 [quant-ph], September 10, 2014. https://arxiv.org/abs/1409.3097

The following methods are applied for this course:

• Presentation and discussion of state-of-the-art research topics.
• Research of relevant literature
• Working in small groups

If you have any questions, please contact care4u@inso.tuwien.ac.at.

The evaluation results from the active participation by giving a presentation about a chosen topic/publication.