The grand challenge in modelling environmental and complex physical systems lies in navigating the Information Bottleneck: how can we compress vast, 4D spatiotemporal dynamics into the representations suitable for Machine Learning (ML) models without sacrificing critical predictive power? This seminar addresses this fundamental trade-off. We will discuss the strategies for Encoding high-resolution data (e.g., using autoencoders or local feature extraction) and assess methods for principled Dimensionality Reduction (such as advanced PCA or manifold learning) to combat the curse of dimensionality. Crucially, the talk will focus on moving beyond black-box predictions by emphasizing model Interpretation and leveraging Physics Understanding to ensure the compressed feature space is not only efficient but physically meaningful, leading to models that are both accurate and trustworthy.
| Open to | all |
|---|---|
| Prior registration | not required |
| Organised by | PICAIS - Passau International Centre for Advanced Interdisciplinary Studies |
| Event website | https://www.picais.uni-passau.de/en/news/news-report/public-lecture-representing-4d-dynamics-in-machine-learning |
| Contact organizer of event | david.sladek@uni-passau.de |
