Agenda (2024)

AGENDA (2024)

Morning Session (9 am – 1 pm CET)
(click to expand the full schedule)
09:00 amSESSION 1
“Don‘t Forget Clinical Reality – How Data-Driven Rehabilitation Can Enhance Therapy Today” by Kim Kristin Peper (TUM MIRMI)

The targeted use of data already has the potential to significantly improve rehabilitation for individual patients today. Through data analysis, insights can be gained to provide real-time feedback for doctors, therapists, and patients during the rehabilitation process. A particular focus is placed on the use of mechatronic aids, especially robots, which can make a significant contribution to rehabilitation. 
However, the integration of data into rehabilitation comes with current problems and everyday challenges. Engineers and data scientists are faced with the task of seamlessly integrating data into the rehabilitation process without neglecting data privacy and security aspects. The role of mechatronic aids is particularly emphasized here, as their integration into existing clinical workflows requires careful planning. 
“The Truth about Data Science in Real Estate: From a Classical Use Case to the Everyday Challenges” by Dr. Federica Fusco (bulwiengesa)

It doesn’t happen every day to read about data science in the real estate industry. During this talk I will present the model, which hides behind “RIWIS Prospect”, one of the tools in the bulwiengesa developed application RIWIS. This tool uses a log-linear generalized additive model to predict rental and purchase prices of apartments and houses in Germany. To train the model, we use millions of data points, which have been collected since 2009. Based on this use case, I will discuss which data we can rely on and then give insights into the challenges and limitations we face in the real estate industry. These include availability of data at different geographical scales, as well as the acceptance of the results and methods by the stakeholders. 
11:00 amSESSION 2
“Applying Causal Models for the Safety Analysis of Automated Transport Systems” by Lina Putze (DLR)

The increasing degree of automation in classical transport systems promises many enhancements such as an increase in performance or improved resource economy. Safety is often mentioned as a main selling point of highly automated systems as the influence of the human operator as one of the most common risk factors is minimized. However, the introduction of highly automated systems entails new classes of hazards that are not relevant for the human operator and thus not represented in current accident data bases. To identify and prevent such hazards during the development phase, a causal understanding of the system and its operational design domain is required. In this talk, we explore how the framework of causal theory can be applied to support the safety analysis process of highly automated transport systems including the identification of appropriate risk mitigation measures.
“An Introduction to Edge AI for Data Scientists” by Ann-Christin Bette (Infineon)

Edge AI is becoming an increasingly important focus area and refers to machine learning models running on local devices. This talk introduces the field of Edge AI for data scientists from a hardware manufacturer’s point of view and explains its fundamental differences from cloud-based AI. Selected real-world examples of Edge AI applications are presented. In addition, the process of developing Edge AI applications, including data collection, model training, and deployment to local devices, is explained. The challenges associated with Edge AI, such as limited processing power and memory are discussed in detail, as well as the best practices and techniques to overcome these challenges. In conclusion, this talk aims to provide data scientists with an understanding of the unique challenges and opportunities of Edge AI.
Afternoon Session (1:15 pm – 5 pm CET)
(click to expand the full schedule)
01:15 pmSESSION 3
“AutoML: Streamlining Machine Learning” by Dr. Katharina Eggensperger (University of Tübingen)

Machine learning (ML) can be frustrating and time-consuming since small design decisions can heavily impact performance. Automated Machine Learning (AutoML) aims to facilitate the development of ML solutions by efficiently obtaining high-performing predictive models via hyperparameter optimization, model selection, and neural architecture search methods. My talk will focus on AutoML methods to design predictive pipelines automatically given a dataset and metric. Join me as we explore the potential of AutoML to navigate the landscape of machine learning model development.
03:30 pmSESSION 4
“Evaluating Ethics and Safety of Generative AI” by Laura Weidinger (Google DeepMind)

How do we know when an AI system is “safe”? In this talk, I review ethical and social risks from generative AI and introduce a sociotechnical approach to evaluating the safety of these systems.  
The landscape of AI safety evaluation is rapidly evolving and of increasing interest to model developers and to the public. We know that generative AI – such as large language models – can produce significant safety concerns, ranging from misinformation to discriminatory bias and malicious uses. In this talk, I review the landscape of these risks and provide a structured and principled approach toward measuring them. Methods at multiple levels of analyses are required to get a full picture of AI safety: safety can be assessed at the level of the technical artifact, requiring disciplines such as computer science and data science; at the point of use, leveraging human-AI-interaction and psychology; and at the point of wide-scale deployment of a new technology, using social science methods. At each of these levels, methodological challenges arise that require interdisciplinary solutions. I point at emerging research that integrates evaluation across these levels and provides a more comprehensive understanding of the safety of an AI system. 
“Retrieval-Augmented Generation: Unlocking The Potential of your Text Data with Large Language Models” by Miriam Kuemmel (Deepset AI)

Large language models have become ubiquitous. While their out-of-the-box performance is impressive, they have some known limitations. An LLM’s world knowledge is bound to its cut-off date, and it was most definitely not trained on the data you work with. But, what if you do want to leverage one for your in-house text data? Fine-tuning LLMs is both skill-intensive and extremely costly, and often doesn’t lead to the desired results. Enter retrieval-augmented generation (RAG): a fusion of LLMs with document search, creating robust generative NLP systems for your text data. In this talk, I will discuss how to construct efficient RAG pipelines and explore strategies to mitigate potential issues, such as hallucination effects. 
05:00 pm – open endGet-together
No program, just delicious food and plenty of networking opportunity.