CAIML #30

CAIML #30 is going to happen on May 14, 2024, at inovex. We will have two talks with additional time for networking.

Talk 1: Robin Senge - Head of Machine Learning at inovex: Uncertainty Quantification: Methods and Applications

Predictions of machine learning models are influenced by different kinds of uncertainty. These arise from lack of knowledge, insufficient data or incorrect assumptions as well as a potentially bad choice of the model class. Uncertainty quantification resembles a set of methods aiming at measuring uncertainties involved in order to be able to mitigate or at least deal with the consequences. This talk introduces methods and their applications. Dr. Robin Senge works as Head of Machine Learning at inovex. As a specialist in machine learning, he designs and implements data-driven use cases applied in commerce and supply chain applications. His research aims at quantifying uncertainty in AI systems as well as improving their interaction with humans through human-centered explainable AI.

Talk 2: Sven Giesselbach - Team Lead Natural Language Understanding at Fraunhofer IAIS: The Potential of Foundation Model Agents and the Need for Custom Small Language Models: A Case Study in the Emergency Trauma Room

Foundation Models have significantly transformed the landscape of data science, particularly in Natural Language Understanding, by leveraging their transfer learning capabilities and agent-based workflows. These advancements enable the rapid development of solutions with minimal or no initial training data. This presentation will delve into a case study within an emergency trauma room setting, where large language models and agents were employed to provide medical staff with essential information and insights in real time, while also automating the documentation process for treatments. The field of medicine, characterized by challenges in data accessibility and availability, stands to benefit immensely from the capabilities of foundation models. Nonetheless, the sensitive nature of medical data and stringent privacy regulations pose obstacles to the deployment of cloud-based models. To address these concerns, we are in the process of fine-tuning a 7-billion parameter Mistral model on German medical texts and instructions, with plans for further customization for tool usage.

See you in May 🤖

Big thanks to inovex for supporting this event! 🙏 Stay tuned for more details and sign up now