DeepLearn 2025: Advanced training, exchange and current developments in the field of deep learning
1 August 2025
From July 21 to 25, Falk Maoro and Benjamin Vehmeyer took part in this year's DeepLearn 2025, an international training event focusing on deep learning, which took place in Maia near Porto in Portugal. The conference was aimed at a broad audience, including doctoral students, postdocs and AI-related industry experts. Current developments in the field of machine learning were discussed in an open atmosphere.
In addition to the varied lecture program, the main focus was on professional exchange. In discussions with researchers and development managers from all over the world, participants were able to gain new perspectives and reflect on their own approaches. Two lecture series were particularly interesting. They were designed to be both theoretically sound and application-oriented.
„Explainability in Machine Learning“
This three-part lecture series provided a structured overview of explainable artificial intelligence (XAI) methods. The first part dealt with classical approaches such as feature attribution and intrinsically interpretable models. In the second part, the explainability of large language models (LLMs) was discussed, for example with the help of probing techniques and mechanistic interpretability. Finally, an insight was given into current approaches to explainability in reinforcement learning, in particular for the comprehensible representation of decision-making processes.
„From Prototype to Production: Evaluation Strategies for Agentic Applications“
This session explored the question of how generative AI systems can be validly evaluated in real application environments. Both technical and methodological aspects were addressed: from the definition of suitable metrics and heuristic procedures to the use of LLMs as an evaluation instance. One focus was on the question of how systems can be iteratively improved through human feedback and how their reliability in productive operation can be ensured in the long term.
Overall, DeepLearn 2025 provided a good overview of current research topics in the field of deep learning. The open exchange with young researchers from different institutions made participation particularly worthwhile for both of them.
Source: Falk Maoro / Benjamin Vehmeyer (FI CODE)
