It's raining data
Josie, your work is about simplifying problems. Can you tell us a bit more about that?
My research focuses on model reduction for Bayesian inference. Essentially, this involves developing smaller, lower-dimensional approximations that can be computed more efficiently from complex, high-dimensional problems. We aim to ensure that these approximations are of the highest possible quality by demonstrating and numerically implementing them. This process involves a significant amount of programming and code development.
Can you give an example for such a problem?
Consider weather forecasting: Every day, a vast amount of data is collected, beginning with an initial estimate of the weather conditions. Throughout the day, additional data is gathered to refine this initial estimate, making it more accurate to the actual conditions. Predictions for the following day are then made based on this updated information. However, the data collection and calculations are very complicated and time-consuming because of the many dimensions involved. In reality, most of the significant dynamics occur on a much smaller scale. Our goal is to identify these smaller-scale structures to streamline and accelerate the calculations.
And how do you find these smaller structures?
So, to recognize and address these issues, my research operates at a theoretical and conceptual level, focusing on basic research. Currently, we are examining the underlying weather models and working on reducing their complexity by accounting for measurement errors. While existing algorithms handle dynamic systems without considering uncertainties, we aim to adapt these concepts to incorporate the effects of uncertainties and distributions, both theoretically and numerically.
“I've found that understanding the importance of research data management from the start has been incredibly helpful.”
Josie König
Josie König's work is about simplifying problems—nothing simple about that. (© Josie König)
Do you make your data accessible to others?
Currently, I am primarily focused on prototyping, so I don't have any real data of my own yet. Instead, I use models and datasets from the mathematical and model reduction communities. Many of these projects aim to ensure the reproducibility and comparability of algorithms, and we utilize their datasets or examples for our work. Additionally, we reference and compare examples from previous research papers. While we don't generate our own data, we make our software publicly available through Git repositories. This allows reviewers and readers to access our work and ensures that the algorithms we use are citable and publicly accessible. We've also collaborated with other authors to encourage the publication of their data and software, which has been very successful.
How do you ensure that reusing research data and software leads to new insights rather than simply replicating previous work?
We first replicate existing algorithms to identify and fix errors, which has already led to valuable collaborations. We benefit from published software, correcting issues together with the original authors. Additionally, by combining existing software and algorithms, we save time and focus on developing new concepts, while ensuring proper citation and facilitating further research through established examples.
Have you ever experienced your software or research being cited?
Not that our work was cited in other papers, but we had very positive experiences in the review process. Reviewers accessed our public repository, conducted their own experiments, and suggested additions based on their findings. They appreciated that our software was available for direct verification and testing on other examples.
Besides manually writing metadata schemas, do you use any other tools to support research data or software management?
I also benefit from being part of a specialized research center focused on data management and reproducibility. We have regular workshops on best practices and tools for code management and archiving. Our center provides support through personal consultations and expert reviews of our repositories and documentation, ensuring clarity and completeness, we have packaging experts who review my Git repository, including the README and my paper, to ensure the code is well-commented and the documentation is clear and comprehensive.
So, it's not just about providing the tools you need, but also about having people who offer real support?
Based on my own experience, I've found that understanding the importance of research data management from the start has been incredibly helpful. In comparison to previous projects or work at other universities, having supportive people available when you don’t know much about it has made a significant difference.
What were the biggest challenges and frustrations you faced when trying to improve research data and software management?
In my community, extensive use of MATLAB is required. The biggest challenges were dealing with this software because it is not open source. Without support from the Collaborative Research Center, it would have been difficult to ensure that reviewers could use our work effectively. Such support is not yet standard at many institutes, though it’s increasingly recognized as beneficial. Publishing code and making it open source helped us, but more concrete incentives could encourage others to share their data and code. Unfortunately, older foundational work often still requires rewriting.
Besides support units and researchers, who else could help in this matter?
I believe professors and supervisors could be more supportive. My supervising professor Melina Freitag was very helpful, but it was within my area of expertise, and I had prior experience in software development. If others lack motivation or awareness, it’s challenging without supervisory backing. Financial support is also crucial. In the math community, there are independent projects that support good documentation, which is beneficial for setting up benchmarks and addressing specific needs. This support is practical, especially when different types of data and needs are involved.
What advice would you give to someone just starting to work with research data?
Honestly, it went well because I had two years of experience as a student trainee in software development. Without that, it would have been much harder. I’d definitely recommend gaining similar practical experience outside of university, as it provided me with valuable tools and skills that I use now.
Frage 1
Which type of support would help you most as a researcher?
Explanation:
She explicitly mentions regular workshops at her research center focused on best practices for code management and archiving.
“We have regular workshops on best practices and tools for code management and archiving.”
She highlights direct expert support as extremely valuable.
“Our center provides support through personal consultations and expert reviews of our repositories and documentation.”
She says financial support is crucial and helps with documentation, benchmarks, and practical needs.
“Financial support is also crucial… independent projects that support good documentation… beneficial for setting up benchmarks.”
She clearly stresses that supervisors and professors play an important role in motivating and enabling good data/software practices.
“I believe professors and supervisors could be more supportive… without supervisory backing it’s challenging.”
Auswertung
richtig beantwortet!
Josie König
diving into the storm of data with her work, using Bayesian methods. In a world where weather forecasting and high-dimensional problems collide, Josie is not just simplifying complex issues but also transforming chaotic data into clear insights. She explains how to navigate this turbulent data and highlights why having the right support is as important as having the right tools.
About the project
The project “Model order reduction for Bayesian inference” aims to make Bayesian inference, an advanced statistical method for making predictions based on uncertain data, faster and more efficient. While this method provides valuable insights, it can be intensive with complex problems. By employing model order reduction, the project seeks to simplify these models without losing essential information, allowing for quicker calculations. The goal is to develop new algorithms that enhance the practicality of Bayesian inference in real-world applications, helping researchers and professionals make better decisions based on data.
Frage 1
Why do you think publishing code openly matters most?
She repeatedly connects open code to reproducibility and comparability of algorithms.
“Many of these projects aim to ensure the reproducibility and comparability of algorithms.”
She describes fixing errors together with original authors and forming collaborations through shared software.
“This has already led to valuable collaborations… correcting issues together with the original authors.”
She gives a concrete example of reviewers using her public repository to test and verify results.
“Reviewers accessed our public repository, conducted their own experiments…”
This is false because she explicitly says benchmarks and documentation are beneficial and supported by funding.
Auswertung
richtig beantwortet!
