Women in Data Science (WiDS) Conference, FAU Erlangen-Nürnberg, 2024

Speaker: Prof. Dr. Stefanie Scholz, Professor for Data Science in Social Economy

Keynote Talk: Enhancing Survey Efficiency through Natural Language Processing

Abstract: Surveys represent a cornerstone of market research, traditionally conducted through in-person or telephonic interactions, which have increasingly transitioned to the digital domain. Yet, while online surveys offer expediency and cost-effectiveness, they often encounter challenges regarding data reliability due to factors such as fluctuating participant engagement, low motivation, distractions, dropouts etc. In that presentation, a method for streamlining the survey process in an “automated” fashion will be introduced. Grounded in the concept of User-generated Content (UGC), this approach involves a structured procedure for UGC retrieval. Leveraging advanced techniques like Natural Language Processing (NLP) and expansive language models, user-generated responses to survey questions (items) are identified and assessed with a heightened degree of accuracy. This methodology affords advantages, including higher reliability due to authentic content, the ability to replicate surveys at any time, facilitate item expansion, and simulate past survey scenarios.

Speaker: Melanie B. Sigl, Managing Consultant for ML

Industry Talk: How to deploy your ML project to production with MLOps

Abstract: Despite the promising outcomes from the proof-of-concept phase, a significant number of machine learning projects terminate prematurely, failing to make the transition to production. This phenomenon is commonly attributed to a myriad of challenges. Above all, financial considerations play a notable role, primarily due to the substantial workload necessary for the transition of a prototype to a production-ready model. In this talk we introduce MLOps, its properties and components, and determine what needs to be done to lift and shift any prototype to a production environment of your choice. We define a simple and easy to follow three step workflow and identify potential problems in each step. To guarantee production deployment, we discuss potential solutions to overcome these problems.

Speaker: Dr. Larysa Visengeriyeva, Head of Data and AI

Industry Talk: Business-aligned AI Products

Abstract: Developing innovative software products starts with an understanding that AI/ML should solve a real problem. However, the whole process, from identifying the use case to the introduction of ML models in the company, is not trivial. Larysa will talk about EventStorming and the ML Design Canvas, which help domain experts and developers get a shared understanding of a business domain and identify use cases for AI/ML technologies. Using the ML Design Canvas, we formulate each use case as an ML problem.

Speaker: Ridhima Garg, Large Language Model Student

Industry Talk: Automatic Prompt Engineering

Abstract: In recent years, the development of natural language processing (NLP) models, particularly large-scale language models, has led to significant advancements in various applications such as text generation, question answering, and language understanding. However, the effectiveness of these models often depends on the quality of prompts used during training and inference. There are recent literature in which large language models (LLMs) have displayed impressive capabilities in generating high quality prompts based upon the base prompt. Automatic Prompt Engineering (APE) is optimizing the prompt in order to elicit the best possible performance. It generates instructions for a task that is specified via output demonstrations. The process is framed as a black-box optimization problem, enhancing the efficacy of prompt-driven tasks.

Speaker: Dr. Elisabeth Preuhs, Project Lead/Scientist

Industry Talk: Artificial intelligence for transforming image-guided therapy

Abstract: Artificial intelligence (AI) is a pivotal driver in advancing image-guided and minimally-invasive procedures for the future. It plays a vital role in processing the complex information flow gathered throughout the treatment path, from pre-procedural data to real-time insights during operations. By leveraging AI and Deep Learning the foundation for improved personalized and precise medicine is shaped. In this talk examples of these transformative projects and building blocks from the predevelopment are highlighted, showcasing their potential to deliver optimal outcomes for every patient.

Speaker: Verena Barth, Co-Founder

Industry Talk: Clarity in complexity: Explainable AI (XAI) explained

Abstract: Machine learning (ML) models are already part of our everyday lives and are behind the scenes used for predictions, data-driven decisions and content creation. Due to their inherent complexity, they are mostly incomprehensible to humans and their decisions cannot be understood. The lack of transparency of these “Black Boxes” is a major drawback and an obstacle to their use, especially in critical areas like the medical sector or the military where interpretability is crucial. In Amazon’s AI-supported application process, for example, women were disadvantaged due to their gender, as the AI was trained on the previously male-dominated employee base. Explainable Artificial Intelligence (XAI) uses various methods to address the problem of lack of transparency in ML models and make the results of the solution understandable to humans. This talk will explore the potential of XAI and show how to promote the interpretability of ML models and trust in them without limiting their (learning) performance. Different XAI methods will be presented and categorized. In addition, the criteria to be considered when selecting and evaluating the suitability of XAI methods will be explained. The talk will end with general recommendations for ensuring interpretable ML models within a development and deployment workflow.

Speaker: Daniela Bernhard, Ph.D. Student

Academic Talk: Optimization with uncertain data via a special stochastic gradient algorithm

Abstract: Input parameters of optimization problems are often affected by uncertainties. For example, the parameters can result from empirical observations or measurements. Concerning the inexactness of these data generating methods, the optimization should include the presence of uncertainties in the data. A common way to model uncertainties in the constraints are chance constraints. Instead of requiring a strict condition for the solution of the optimization problem, chance constraints relax the constraint to only hold with a given probability less than one. In this talk, we consider joint chance-constrained problems with continuous distributions. To solve them, we apply an improved version of the Stochastic Gradient Descent algorithm, namely the Continuous Stochastic Gradient method, that has recently been introduced in the paper “CSG: A new stochastic gradient method for the efficient solution of structural optimization problems with infinitely many states” by Pflug et al.. Therefore, we first reformulate and approximate our uncertain problem with smoothing and penalization techniques to apply the gradient method. Under very mild assumptions, our approach is applicable to a wide range of joint chance-constrained problems. We illustrate the efficiency of our approach with an application to gas networks. The numerical experiments demonstrate that the approach quickly finds nearly feasible solutions for this difficult class of problems, even for realistically-sized instances. This is joint work with Frauke Liers and Michael Stingl (both FAU).

Speaker: Shofiyati Nur Karimah, Consultant Data Governance

Industry Talk: Automatic Engagement Estimation for Distance Education

Abstract: Engagement is an essential component of the learning processes associated with positive learning outcomes. Since the learning paradigm has shifted to enable more distance education practices, machine learning-based automatic engagement estimation methods have been proposed as a new way to measure learner engagement in distance education. Nevertheless, implementing automatic engagement estimation should ensure technological and ethical impact responsibilities. This talk will explore a design principle for an end-to-end integration of automatic engagement to measure learners’ emotional engagement in a synchronous distance education practice.

Speaker:  Susanna Braun, Cybersecurity Professional at Siemens Healthineers

Industry Talk: Operational Technology (OT) Security

Abstract: In today’s interconnected world, securing our critical infrastructure is crucial. This speech focuses on Operational Technology (OT) security, vital for industries like energy and manufacturing. It highlights the unique challenges of OT environments, such as outdated systems and IT-OT convergence, through real-life examples like the Colonial Pipeline attack. Attendees will discover practical strategies for safeguarding these essential systems, including risk management, network segmentation, and compliance with industry standards. Join this call to understand why OT security is a critical step towards ensuring our future safety and stability.

Speaker: Prof. Dr. Anne Koelewijn, FAU, Chair of Autonomous Systems and Mechatronics

Academic Talk: Understanding Human Movement through Machine Learning

Abstract: Measuring human movement is challenging, since invasive sensors can only be used in exceptional cases. Therefore, we have to rely on what we can see and measure on the outside. This information is very accessible, since many datasets exist that contain human movement recordings. Therefore, it is our goal to try and extract as much information as possible from these available datasets using machine learning methods. In this talk, I will discuss how we have used machine learning to enhance our understanding of human movement. Specifically, I will delve into examples where we used machine learning models to estimate of movement variables “in the wild” using sensors and methods that can be applied outside of a laboratory, and to improve our knowledge on muscle mechanics and energetics.