[AlgoEarth] Learning from limited data in Earth system sciences: optimal selection of datasets and methods

We kindly invite you to join us on November 5-7 to a workshop taking part within the semester of Algorithmic Earth System Sciences (AlgoEarth) at the University of Cologne (Please use the entrance at Höninger Weg 100, 50969 Cologne. Ring the bell labeled “Hörsaal Universität Köln – Öffnungszeit 7–18 Uhr” to enter; the door will open automatically. The session takes place in Lecture Hall 4.001 on the 4th floor.):

Learning from limited data in Earth system sciences: optimal selection of datasets and methods

The workshop will take place at the University of Cologne. The goal is to bring together experts in methods development and in Earth System applications to tackle specific problems arising in learning tasks when the data availability is limited. Please see the text below for more detailed information and invited contributions.

More information is provided here:  https://imfess.uni-koeln.de/algoearth/events/workshop-learning-from-limited-data-in-earth-system-sciences#c175402

Please register until September 15 under this link (the registration is free of charge): https://survey.uni-koeln.de/index.php/921438?lang=en

Learning from limited data in Earth system sciences: optimal selection of datasets and methods
Small data challenges are particularly common in Earth System Sciences. Examples include the study of rare and extreme events, which tend to have high societal impact but are often poorly forecasted by classical models. The workshop will discuss targeted machine learning methods for “small data” problems – when the underlying learning task is highly underdetermined, due to a large problem dimension relative to the size of training datasets.  The goal is to bring together experts of data and computer sciences with experts of Earth system sciences that use and possibly develop such methods. We will discuss and identify:

• Optimal selection or treatment of data: How can one augment the information content of the limited dataset by smart data selection or preprocessing approaches? How can one benefit from multiple data sources?
• Highly efficient approaches to the small data problem, as well as overfitting challenges in artificial intelligence
• Techniques to study and improve reliability of machine learning methods and quantify uncertainty.
• Key challenges motivated by AI application in Earth system sciences.

Invited Speakers:

• Claudia Acquistapace
• Davide Bassetti
• Dwaipayan Chatterjee
• Francesco Marra
• Julian Quinting
• Nedjeljka Zagar

We welcome contributions related to:

• Methods that can learn efficiently from little data
• Optimizing the information content from limited data (e.g. for extreme events such as hail or rain floods, using alternative data sources, fine-tuning methodologies, etc.)
• Optimizing loss/fitness functions for rare events, e.g. reweighing
• Methods that can provide guarantees on false positive discovery rates
• Methods that disentangle epistemic and aleatoric uncertainty
• Application-specific interpretability methods

For further information about the AlgoEarth Semester, please visit our website.