On May 27th, Prof. Dr. Stefanie Jegelka from the Technical University of Munich (TUM) and Massachusetts Institute of Technology (MIT) held a talk “Learning to Reason with Graphs”. The talk was part of the IUC Applied Research Talks series and was co-hosted by the Predictive Town Hall Meetings, it took place in the Auditorium and virtually for TUM and SAP experts and employees.
The talk focused on how Graph Neural Networks (GNNs) and Large Language Models (LLMs) can support algorithmic reasoning over graph-structured data.
Sharing that GNNs align well with dynamic programming, illustrating the broader concept of algorithmic alignment, where the architecture of the neural network is aligned with the algorithmic structure of the target function or task. This alignment enables neural networks to learn faster, perform better, and be more stable when dealing with new or different data. Further addressed that small LLMs, when fine-tuned using reinforcement learning, can learn to reason over graphs even without human-labelled data. These models not only improve significantly in performance but also transfer well across different graph tasks, datasets, and input formats.
After the captivating talk, participants were able to ask questions and engage in a discussion round accompanied by food and refreshments.