September 10-14, 2023, Zugspitze, Germany
Information Theory as a Bridge Across the Geosciences and Modeling Sciences
What this workshop is about
1. Overview: Combining knowledge from theory and learning-from-data is currently seen as the most promising paradigm for Geoscientific progress. Further, the growing understanding that information is fundamental to how we encode our understanding of reality has rekindled interest in how Information Theoretic principles can be implemented in the Geociences.
Aim: This workshop builds upon a series of successful workshops and summer schools entitled “Information Theory and the Earth Sciences” that we have previously held at the Schneefernerhaus (Germany 2016) and Santander (Spain 2018, 2019), as well as numerous conference sessions at AGU and EGU. It seeks to advance development of a framework that supports Geoscientific investigation through enhanced Extraction, Representation, and Interpretation of Information from data. By exploiting the nexus of Information Theory, the Modeling Sciences(including Machine Learning), & Domain Relevant Theory, we can expect to enhance both the predictive capabilities of domain-relevant models and their suitability as a basis for reasoning and understanding.
Participation & Outcomes: We welcome workshop participants who actively (or aspire to) work across at least two of the following domains – Information Theory, Modeling Science, & Domain Relevant Theory. The expected outcome is enhanced dialogue and collaboration, resulting in progress towards a general framework for constructing domain- and task-relevant models that can learn from data while maintaining representational structures that are consistent with our physical understanding of the world (and are therefore interpretable).
Format: Building upon high-level presentations, substantial time will be devoted to moderated small-group discussions focused around specific pre-announced topics. Each participant will also have the opportunity to contribute a scientific poster.
2. Statement of the Problem: Conventional modeling of dynamical Earth Systems is rooted in an incomplete understanding of the world, as expressed by existing physical-conceptual theories. As such, Physics/Theory–Based (TB) models can represent severely lossy (and arguably even seriously incorrect) compressions of the information content that is extractable from available data. The consequence is that Data-Based (DB) models, rooted in modern Machine Learning (ML), typically outperform TB models for even relatively simple predictive tasks. Combining knowledge from theory and learning-from-data is therefore currently seen as the most promising paradigm for Geoscientific progress, but to date, a consistent framework and language to do so is lacking. In this regard, IT also allows us to approach fields such as Sensitivity Analysis and Uncertainty Characterization/Quantification in new, and potentially more powerful, ways.
3. Focus of the Workshop: The goal of the planned scientific workshop is to advance development of a framework to support Geoscientific investigation through enhanced:
- Info Extraction: Using modern Modeling Science to extract Geoscientific Information from Data.
- Info Representation: Using this Information to construct suitable (task-relevant) Models of dynamical Geoscientific Systems.
- Info Interpretation: Doing so while maintaining Representational Structures that are consistent with our physical understanding of the world and are therefore interpretable.
- Info Generation: Using the derived Repesentations to generate Information about aspects of the Geoscientific Systems that are not easily observable.
- Info-based Reasoning: Developing the abilities for Causal Analysis, Counterfactual Investigation, and Hypothesis Testing and Revision
As such this workshop seeks to exploit the nexus of Information Theory, the Modeling Sciences, & Domain Relevant Theory (see Figure 1). By doing so, we seek to enhance both the predictive capabilities of models and their suitability as a basis for reasoning and understanding.
- Information Theoretic Component: Information Theory can be viewed as the overarching (bridging) component of this workshop, encompassing the statistical Shannon perspective, the computational Algorithmic perspective, and what might be called the Representational perspective (expressing how info is encoded via representational choices). In this regard, the Information Bottleneck principle, the Maximum Entropy principle, the Minimum Description Length principle, and the Physical Nature of Information can be considered central issues.
- Modeling Science Component: Modeling Science is the component by which domain knowledge is encoded and can be viewed as encompassing the Probability and Statistics perspective, the (Deep) Machine Learning perspective, the Computational Science perspective, and the Physics-Based Modeling perspective.
- Domain Science Component: Clearly, the Domain Sciences (here intended to encompass the hydrologic, atmospheric, and ecological sciences, among others) provide the context within which the framework for supporting enhanced scientific understanding is to be viewed. It is through this component that domain-relevant knowledge (including conservation and thermodynamic principles) is to be acquired and injected, and within which enhanced prediction and understanding are sought.
Figure 1: Bridging across Information Theory, Modeling Science & Domain Relevant Theory.
Examples include: SIT = Shannon Info Theory (statistical perspective), AIT = Algorithmic Info Theory (computational perspective), RIT = Representational Info Theory (representational perspective), IB = Information Bottleneck, CS = Computer Science, ODE = Odinary Differential Equations, DL = Deep Learning / Machine Learning, STAT = Probability and Statistics, ATMO = Atmospheric Science, THM = Terrestrial Hydrometeorology, HYD = Hydrological Science, ECO = Ecological Science.
The Mural – our central communication and synthesis platform
We have established a mural for the workshop, which will serve as the central communication platform before, during, and after the workshop.
Access the mural here. Anyone with this link can access the mural as a visitor with view-only rights. For active use of the mural, we will invite all workshop participants as guests to the mural, which then permits viewing&editing. Expect an invitation Email in the next few days with further instructions (involves creating a free account).
We have compiled an overview of the structure and the usage of the mural here (in the mural, you will find it in section 0.0 (“Welcome!”).
Feel free to engage with the mural at any stage. Throughout the workshop, we will leverage the documentation of activities and discussion via the Mural to synthesize workshop outcomes and identify the next avenues together.
In mural section 0.0 (“Welcome!”), you will find all general information on the workshop: schedule, participants list, logistics information, information on the hiking trip, and information on how to use the mural.
If you have further questions, please post them on the mural in the section “Help Desk.” You can, of course, also write an e-mail to one of the workshop organizers.
Looking forward to seeing you at the Schneefernerhaus!
Uwe Ehret, on behalf of the organizing committee:
Uwe Ehret (Karlsruhe Institute of Technology)
Hoshin Gupta (University of Arizona)
Cristina Prieto (IH Cantabria)
Praveen Kumar (University of Illinois, Champaign)
Marvin Höge (EAWAG)
Andrew Bennett (University of Arizona)
Maoya Bassiouni (University of California, Berkeley)