Agenda Machine Learning Week Europe 2024

18 - 19 November 2024, Munich

Today’s LLMs such as ChatGPT show an impressive performance and have the potential to revolutionize our daily life. All these LLMs are based on the Transformer architecture with the Attention mechanism at its core. Due to the quadratic scaling with context length, Attention makes processing of long sequences very expensive. In this talk we present xLSTM, a novel architecture for LLMs that scales only linear in context length while still outperforming Transformers on language modeling.

The talk will present TotalJob’s learnings and approach taken towards productionizing GenAI / LLMs for building an empathetic, accurate and compliant job search companion (available over WhatsApp). It will cover the product decisions in going from human to LLM chat and techniques used to reduce LLM costs by 70%, with hallucinations, bias and toxicity at near zero %. Finally, the talk will give an outlook on the future path around a multi-modal and the multi-lingual future of conversational journeys.

Train drivers have different driving styles which require different amount of energy. These different styles are difficult to analyze to define an energy-efficient driving strategy. We present a reinforcement learning approach that learns to drive long-distance trains more energy-efficient while maintaining punctuality and safety. The method was used by DB Fernverkehr on a test-drive and showed an unexpected driving strategy that was very energy-efficient and material friendly.

Even (and more so) with advances in deep neural networks, understanding data remains a central challenge in data science, but central to its value. Data mapping techniques help us to see salient features of our data sets, explore different dimensions of the underlying systems, and support statistical decision making. I will discuss and demonstrate practical methods to data mapping, from dimensional reduction and clustering over self-organizing maps to deep learning-based variational autoencoders.

An effective picker schedule in warehouses in Quick-Commerce addresses the challenge of resource utilization while maintaining low delivery time. The talk focuses on the challenges in running decision science experiments with humans involved, and embracing learnings by strategizing: How to utilize limited and unclean data caused by stochastic activities and human-error in warehouses? How to draw insights while navigating uncertainties? How to optimize operational decision making processes?

In the age of increasingly complex foundation models, the question arises as to what extent the underlying correlation structures also represent causal effects. In the field of time series analysis, the concept of causality is a central component and there are established techniques for identifying non-causal spurious correlations. The Deep Dive illustrates these concepts and techniques and shows how they can be successfully combined with state-of-the-art ML algorithms. No formulas required!

How do you motivate and guide an existing organization to create and share analytical data products to improve the results of analytics, machine-learning and ai? We will give you insights into our data product cookbook that motivates data mesh and data products, supports data consumers, and provides a step-by-step checklist to build a data product – from business model canvas to data models and lifecycle management.

In this session, we will have an in-depth exploration of chatbot innovation applied in a real-world setting at Endress+Hauser. This session will focus on our practical implementation of Large Language Models, combined with the capabilities of Azure’s Semantic & Vector Search and strategically advance Retrieval-Augmented Generation (RAG) approach, to build a chatbot that is not only highly responsive but also remarkably intuitive .We will provide a comprehensive overview of our approach, detailing the process of integrating technology stacks and navigating the complexities of chatbot design, demonstrating the tangible benefits of these technologies in a real-world, industrial setting.

In many real-world scenarios, like quantitative finance, datasets exhibit a probability distribution with significant skewness. Neglecting fat-tails can lead to increased estimation errors and potential business losses. We introduce a new classifier, termed Elliptical Discriminant, that generalise Quadratic Discriminant Analysis (QDA) to handle fat-tailed data. This is demonstrated using a multivariate dataset with fat-tails and a comparative study with other well-known classifiers.

A roundtable on Generative AI’s impact on Data Engineers is crucial to discuss AI’s capabilities, skills evolution, ethical considerations, and innovation in data management. It’ll address AI’s role in augmenting vs. replacing human tasks, necessary workforce adaptations, ethical deployment, and policy development, ensuring a balanced technological advancement that benefits society and the workforce.

For this table discussion, we want to tackle the following three main questions:

1. How can organizations successfully implement a Data Mesh architecture, and what challenges might they face during this transition?
2. What are the key components of a successful Data Product, and how can organizations ensure they deliver real value to stakeholders?
3. How does the adoption of Data Mesh and Data Products impact organizational culture and team dynamics?

Navigating the complexities of airline claims can be a challenge for both agents and passengers. Claim Summarizer, a novel tool powered by Generative AI (Gen AI), tackles this challenge by providing claim agents with concise, AI-generated summaries of complex claims documents. This presentation explores how Claim Summarizer empowers agents to make faster, informed decisions by highlighting key details and extracting vital information. Moreover, the Claim Summarizer is our blueprint of future LLM Ops. By implementing the Claim Summarizer, we have established robust LLM Ops processes that can be used in our new products and accelerate the time to market of new Gen AI innovation in the Lufthansa.

Special measures must be provided for a highly automated production of safety critical products. Therefore, a system focused on AI-based visual inspections is developed. This should improve the performance, stability and transparency of the particular inspections, the scalability of inspections across assembly lines and especially the quality assurance. The system architecture and the actual implementation for bringing state of the art neural networks into the production will be presented.

Companies implementing AI are caught between prospects of success and societal concerns. Regulations attempt to mitigate the insecurities that arise through the widespread adoption of AI. This session helps corporations struggling to operationalise these standards by providing a step-by-step guide to embed trustworthy AI in their organisations. With a brief foray into strategy, processes & organisational design this talk will show ways to improve the ethical footprint of your company.

The last two years have been wild! LLM breakthroughs, massive awareness and LLM demos everywhere. On the other side, not many made it beyond the POC stage. In 2024, we saw teams on two different trajectories: Some get lost in the POC stage. Some go to production. In this talk, we increase your odds of being on the latter. We walk through key technologies, assess their maturity and in which cases they are the “right hammer”. We share learnings from our work with thousands of Haystack open source users.

Data quality plays a pivotal role in AI, especially when dealing with consumer data. Join us as we delve into entity resolution for consumer data, blending modern machine learning and graph-based algorithms with human expertise, to balance automation, high data quality and manual corrections. We will provide a comprehensive overview of our data pipeline, addressing the potential and limitations of vector search, and detailing the challenges of ensuring data quality for billions of records on a daily basis.

At KSB SE & Co. KGaA, the number of productive ML-models increased with the growing customer base of its digital product KSB Guard. As a result, the need for a clear separation of responsibilities between data scientists and operations increased. In particular, the further development of already productive models became a pain point. In this talk I want to present a service architecture that promotes seamless collaboration between data science and operations teams.

Machine learning models should not live in a vacuum. When a predictive model is exposed to the user to enrich a scientific or industrial workflow, incorporating human feedback is essential for the model to improve over time, overcome distribution shifts, and learn novel phenomena. We address the engineering challenges of building interactive active learning systems on a practical example of large-scale video analysis.

Psychological AI refers to machine learning models that originate from or are attuned to human information processing. These powerful models are simple and transparent and have been shown to outperform more complex, data-intensive models in predictive accuracy. Over the last 1,5 years, we have been working with the Schufa to develop such a model and make the Schufa-score fully transparent – a global first in credit soring. This talk describes our journey.

Model-based process characterisation is key for implementing real time release testing within pharmaceutical manufacturing. Andre illustrates the integration of data-driven modelling and machine learning to craft resilient release strategies capable of accurately predicting the products’ quality attributes. These strategies result in optimised lead times in the supply chain and foster a comprehensive and timely understanding of the process performance, thereby generating substantial value.

RLHF is critical method in NLP which learn from human interactions, generating more accurate and contextually appropriate responses. In this session, we will explore what is RLHF. We will start with discussing underlying principles behind it, various techniques used and finally a real-world example including data collection, reward design, and model evaluation to illustrate its effectiveness. By the end we will discuss its limitations, how it can be adapted to different use cases and domains.

The attendees will explore how Swiss Mobiliar processes insurance quote recommendations faster and more accurately by unlocking hidden data patterns with NLP-based AI functions embedded in SQL following the paradigm “Bring the algorithms to the data – not the other way around”. An example from the world of soccer will be followed by a discussion of the objectives and results of the app for quote recommendations. We will also discuss accuracy, traceability and trustworthiness of the predictions.

Of all the data projects that fail, 80% fail due to poor communication. When decision makers don’t understand a data product, they don’t need more information: they need less of it. In this talk, Jack and Julia will show how data storytelling is helping to promote data products and get buy-in at a pharma company. Tune in to learn the process of shaping the narrative and finding a data product marketing strategy. Get expert tips on how you can use data storytelling for your own data / AI product.

Keyword based search engines will fail to support random queries like ‘‘red shirt with black collar and avenger cartoon in front” in case the product description has not got these details. In Fashion e-commerce such queries are becoming popular. To support this, we can query product images directly by converting images into multimodal embeddings and then searching on them. This talk will discuss the approach and show how to modify the keyword-based search engine to support such queries.

Migration, inflation, climate change, and more present us with numerous uncertainties when making socio-political or strategic business decisions. How can we still make robust decisions? In our BMBF-funded project, InnoTwin, we are developing a digital twin of German society. This allows us to conduct simulations under various conditions to identify actions that yield robust outcomes across a wide range of scenarios.

For healthcare professionals, treating patients in the best way possible is a challenge in more ways than one. Extensive therapy guidelines – though necessary – contribute to the complexity of their working life. To support treatment decisions, Campana & Schott, together with Charité, developed a smart assistant which recommends treatment according to said guidelines based on the patient’s diagnosis and medical history. We will present you our results and envision potential further applications.

Duplicate records can have a negative impact on many areas of a business. Current methods to detect duplicate records use traditional NLP techniques known as “Entity Matching”. An improvement to this traditional method can be achieved by incorporating GenAI techniques that do not entail any calls to OpenAI. Not only does this produce better matches, but it also keeps the data safe, since no information is transferred externally.

The Table Captains present the three key take-aways from their discussion round.