Annika Stein

Research: jet tagging techniques (applying deep learning, ideas from AI safety). With the help of latest architectures like transformer models, I am developing a new generation of algorithms which exploit low-level features in a unique way. My particular interest is targeted at improving not only performance, but also at creating robust taggers that can withstand input distortions and which generalize better to experimental data, being trained on (augmented) simulation only. For Run 3 of the Large Hadron Collider, CMS will use the first (adversarially) robust tagger in production, and I am looking forward to analyses profiting from this development. Studies so far are summarized here. Further activities related to this topic, in an interdisciplinary way, are now coordinated on GitHub. The tools under investigation play a crucial role when studying the properties of the Higgs boson, for example when probing the coupling to charm quarks. I set up a first implementation of the VHcc Zll channel in a pure python framework which can be used for further studies and improvements, compared to the Run 2 version. Became the first one to study data/MC agreement during the ML development process, ergo the only one who performs the whole chain in one go as part of the dev cycle. Certain investigations are unique to my thesis and can only be repeated if you use the full development framework that I introduced for this purpose, but are not part of typical commissioning workflows that don't include the machine learning aspects. On the other hand, usually the software developers focus on the software and simulation performance only, while I also cared about how the algorithm works on detector data. I look forward to apply this paradigm in my new collaboration, such that no surprises await upon introducing a new generation of taggers. I created a new adversarial technique that effectively solves the problems of its predecessor, and because the employed strategies are universally interesting, I engage in preparation of low-level quantities for machine learning purposes as found in Open Data.

Service work for CMS: extending a specialized data tier / framework to study jet tagging algorithms with additional variables that are relevant for commissioning. Code reviews and contributions for commissioning, efficiently hunting 🐛 in the code. Providing a new skim for monitoring purposes of b-tagging observables at HLT. Implementation of adversarial training for next generation of flavour tagging algorithms (used starting from Run3). Building a new software framework that allows training and evaluation in one go.
BTV RECO contact, reviewing pull requests on GitHub, creating some myself to keep the b-tagging software up-to-date.
Contributions to (selection): CMSSW, DeepJet, DeepNTuples, PFNano, BTVNanoCommissioning, VHcc-cTagSF, CMS Machine Learning Documentation, CoffeaRunner, VHcc.

Teaching and Supervision:

• Co-Supervision of several graduate students on the topic of AI safety / jet tagging / application to physics analysis.
  
  • Interpretability of b-Tagging Neural Networks for the CMS Experiment (Summer Kassem, 2022, Master thesis)
  • CMS Open Data for AI safety studies (Taylor Briggs, 2023, DAAD Rise Germany Internship)
  • AI safety (DeepJet) (Hendrik Schönen, 2023, Master thesis)
  • VHcc Analysis (Valentyn Vaulin, 2023, Master thesis)
• Assistant for the Advanced Laboratory Course (B.Sc.).
• Assistant for the Laboratory Course Particle Physics (M.Sc.).
• Teaching Assistant for Datenverarbeitung in der Physik (B.Sc.), coordinator of tutorials / exercise classes / exams.
• Checked exercises for an introductory course for future physics students ("Studieren Erfahrbar Machen").

For the Research Training Group "Physics of the Heaviest Particles at the LHC":

• Co-Organizer of the Annual Retreat (2022)
• Co-Organizer of the RTG Colloquium (2022 - 2023)

Reviewer for

• EPJ C
• JOSS.

Editing / proofreading of course material for students in the first semester, with online content related to Experimental Physics 1. My tasks included checking the technical correctness (physics), language, layout and formatting using MathJax (a LaTeX derivative for JavaScript).

Tutor for an exercise class in Experimental Physics 1 and 2. My tasks included grading student's course assignments, as well as providing a weekly tutorial session and supervising the exam.

Tutor for an exercise class in Experimental Physics 2. My tasks included grading student's course assignments, as well as providing a weekly tutorial session. At the end of the semester, I supervised and graded the final exam.

Tutor for an exercise class in Mathematical Methods (B. Sc.). My tasks included grading student's course assignments, as well as providing a weekly tutorial session. At the end of the semester, I supervised and graded the final exam.

Tutor for the "Lernwerkstatt Mathematik", i.e. coaching undergraduate students in mathematics (or other areas in STEM, like physics or computer science). My tasks included helping students with their weekly homework, pointing them to relevant resources (e.g. books, web), giving hints, explaining course material and concepts (examples: proof techniques, problemsolving, learning strategies) or simply supporting students in their first semesters.