Prof. Dr. Felix Naumann

The meeting will be dedicated to organisational issues of our cluster:

-          New members (DSE Research School and member groups)

-          Retrospective of our retreat

-          Weekly meetings

-          Web site and fall reports

-          Shared tasks

-          Other cluster activities

Georg Tennigkeit

Haptically Rendered Impossible Space for VR

Abstract: Two unsolved problems of Virtual Reality are 1) exploration of virtual worlds through natural walking and 2) large-scale haptics. We propose solutions to both problems that complement one another:

To enable natural walking in a small space, we place the entire play area in a grid of 2x3 tiles, which takes up 3m² of physical space. This grid can contain more than six virtual tiles by overlaying multiple tiles in the same position, a technique known as impossible space. As impossible spaces commonly lose the continuity of the larger space, we have defined repeatable walking patterns that can be rationalized into a non-overlapping motion.

Due to the limited room layouts possible in 2x3-Space, haptic walls become more feasible. We place vertical rotary actuators at each pillar and attach an upright wall panel to them. Each wall panel can represent four of the seven possible walls between two tiles. We describe a motion control algorithm that considers the impossibility of the space and seamlessly transitions between haptically rendering a wall and making way for the user.

Linea Schmidt

Extraction of Crohn’s Disease Clinical Phenotypes from Clinical Text Using Natural Language Processing

Abstract: Background: Crohn’s Disease (CD) patient heterogeneity in clinical practice is captured by the Montreal Classification. While the underlying concepts, disease behavior and age at diagnosis, are relevant outcomes and covariates in studies from real-world data, extracting this clinical information through manual chart review is labor-intensive and with limited scalability.
Methods: We developed and evaluated automated phenotyping algorithms to extract disease behavior and age at diagnosis from clinical narrative texts, using a rule-based approach based on the spaCy framework, and an approach based on zero-shot inference. The underlying data included 49,572 clinical notes and 2,204 radiology reports from 584 CD patients of the Mount Sinai Crohn’s and Colitis Registry. A test set of 200 clinical texts
per classification category was labeled at sentence-level, in addition to patient-level ground truth data. The algorithms were evaluated based on their recall, precision, specificity values, and F1-scores.
Results: For the labeled dataset, an overall Cohen’s kappa inter-annotator agreement of 0.84 was achieved. The rule-based approach yielded high recall and precision values (0.75 - 1.00) on a note level for the behavioral disease phenotype using clinical notes, with slightly reduced performance using radiology reports. For age at diagnosis, recall and precision values of 0.81 and 0.88 were achieved on note-level, respectively. For both categories, the performance on patient- compared to note-level was reduced, potentially due to the accumulation of false positives and limitations in the data availability.
Conclusion: Based on our newly annotated dataset, we demonstrated the feasibility of automatically extracting disease behavior and age at diagnosis from clinical text. The resulting labels may facilitate extensive cohort analyses based on electronic health records, and support chart review processes in the future.

Sumit Shekhar

Abstract: Achieving a PhD is a monumental accomplishment, but the path to this achievement can be fraught with stress, potentially impacting the health and well-being of doctoral students. Without timely intervention, these challenges can significantly diminish the quality of life and academic performance. In this talk, I will explore the multifaceted dimensions of a PhD student's health, encompassing the social, mental, and physical aspects. The objective is to provide a set of general guidelines for maintaining one's health while navigating the demanding world of doctoral studies. Moreover, I will share some real-world examples of burnout, equipping attendees with the knowledge to recognize and address such issues. This talk draws from personal experiences and the collective wisdom of fellow PhD students, offering insights to enhance the well-being of those on the challenging path to academic excellence.

Ilin Tolovski

Title: Interoperability in Data Science Pipelines

Abstract: The development of data science pipelines (DSP) has been steadily increasing in the past decade, reaching its peak in 2021. While the growth of applications can be attributed to novel algorithms and analytics libraries, the interoperability of new DSP has been limited. In a comprehensive analysis of over 494k Python repositories available on GitHub, we have found that only 34% are data science repositories that provide access to their input data. The absence of input data leads to projects that are difficult to execute in a new environment, and results that are difficult to reproduce. We propose Stork, a system for automated data and pipeline migration. By performing static code analysis, Stork removes the human from the loop when migrating projects and reproducing results. Stork supports the migration to database management systems, such as PostgesSQL, as well as cloud object storage, AWS S3.

Prof. Dr. Patrick Baudisch will talk about three (of potentially many) stages in the evolution of a research career.

Dr. Jan Buys.

Linguistically-informed Language Modelling for (South) African Languages

Abstract: Advances in Large Language Models have been driven by scaling the amount of training data and the size of models. However for most of the world's languages this approach is not sufficient due to vast discrepancies in the amount of online text available per language. Additionally, multilingual LLMs do not explicitly take differences in linguistic structure between languages into consideration, which can further exacerbate under-performance on low-resource languages. In this talk I will present some of our work on developing language models for South African languages. I will present some datasets we have collected, and show how adapting languages models for a particular language family (the Nguni languages) can improve performance. Then I will focus on an approach called subword segmental language modelling, which learns how to segment words into subwords while being trained for language modelling. This is motivated by the rich morphological structure of many African languages. The subword segmental approach is applied to machine translation and data-to-text generation. The resulting models outperform other approaches in very low-resource settings in particular. Finally I will discuss injecting knowledge from knowledge graphs into language models for South African languages. 

Cavit Özbay

Title: Multi-Signatures for Ad-hoc and Privacy-Preserving Group Signing

Abstract: Multi-signatures allow to combine individual signatures from different signers on the same message into a short aggregated signature. Newer schemes further allow to aggregate the individual public keys, such that the combined signature gets verified against a short aggregated key. This makes them a versatile alternative to threshold or distributed signatures: the aggregated key can serve as group key, and signatures under that key can only be computed with the help of all signers. What makes multi-signatures even more attractive is their simple key management, as users can re-use the same secret key in several and ad-hoc formed groups. In that context, it will be desirable to not sacrifice privacy as soon as keys get re-used and ensure that users are not linkable across groups. In fact, when multi-signatures with key aggregation were proposed, it was claimed that aggregated keys hide the signers’ identities or even the fact that it is a combined key at all. In our work we show that none of the existing multi-signature schemes provide these privacy guarantees when keys get re-used in multiple groups. This is due to the fact that all known schemes deploy deterministic key aggregation. To overcome this limitation, we propose a new variant of multi-signatures with probabilistic yet verifiable key aggregation. We formally define the desirable privacy and unforgeability properties in the presence of key re-use. This also requires to adapt the unforgeability model to the group setting, and ensure that key-reuse does not weaken the expected guarantees. We present a simple BLS-based scheme that securely realizes our strong privacy and security guarantees. We also formalize and investigate the privacy that is possible by deterministic schemes, and prove that existing schemes provide the advertised privacy features as long as one public key remains secret.

Prof. Dr. Tilmann Rabl

Title: How to safely use and properly attribute your own work in your PhD thesis

Abstract: Computer Science PhD students at HPI and elsewhere typically publish their research at conferences and in journals before submitting their PhD theses. It is well accepted practice to then use these papers verbatim in part or completely in the thesis. However, since most papers have more than a single author, it is important to properly attribute the work in the thesis. In this talk and discussion, we will look at the topics authorship, official PhD thesis regulations, common constellations of authorship, and I will present a structure to transparently and safely attribute all authors work in the thesis.

Prof. Felix Naumann with Sebastian Schmidl

Title: Availability, Repeatability, Reproducibility

Abstract: Recently, many computer science communities have started developing artifact review processes to embrace repeatable and reproducible experimental results reported in papers. In this talk, we will establish the need for repeatable experiments in computer science, explore the definitions of availability, repeatability, and reproducibility, check the artifact review processes based on an experience report from the DB community (SIGMOD), and discuss ways that we can support these efforts and thus establish better scientific practices. A good starting point are the artifact review definitions and policies laid out by the ACM: https://www.acm.org/publications/policies/artifact-review-and-badging-current.

Katerina Niklanovits

An Introduction to Twin Width

Abstract: Twin width is a width parameter introduced by Bonnet, Kim, Thomasse and Watrigant [FOCS'20]. Due to its various structural and algorithmic applications it has been studied intensely since then. In this talk we are focusing on understanding the definition of twin width through some examples on how to compute it for simple graph classes. We are also going to see an 2^O(k)n- algorithm developed by Bonnet, Geniet, Kim, Thomasse and Watrigant [ICALP'21] for k-Independent set when a d-contraction sequence is given.

Prof. Dr. Holger Karl

Technical writing: Clarity and Grace

Abstract: Why should I bother to write well? Because clear writing fosters and reflects clear thoughts. And because it increases acceptance chances of your papers.
In this talk, with plenty of exercises for the audience, we go over typical problems in technical writing. Maybe, we even find ideas how to fix them.

Dr. Fabian Panse

Tabular Data Synthesis for Data Management

Abstract: The problem of generating synthetic data is almost as old as modern research itself. However, with the advent of generative AI, new possibilities for synthesizing tabular data have emerged that go far beyond the capabilities of traditional statistical or rule-based approaches. Most of this new research comes from the ML community, where ML models need to be fed with useful training data. Since many data management use cases also require synthetic data, it makes sense to adapt these research results. Nevertheless, those use cases, such as query optimization, have different requirements than ML use cases. Requirements that are currently not met by such modern synthesizers. In this talk, we will give an overview of the current state of the art in the field of tabular data synthesis and discuss open challenges in the context of generating synthetic tabular data for data management.