Congratulations!
Congratulations to Tobias Schanz, who defended successfully his PhD thesis on the 4th June 2024 at the University of Hamburg. Tobias is working at the department Model-Driven Machine Learning. He gives us an insight into his research work on his doctoral thesis with the title Quantifying and Managing Uncertainties Advancing Deep Learning Approaches for Earth System Sciences.
We could also call it: Navigating Uncertainty in Machine Learning for Environmental Science. During my PhD, I embarked on a journey to bridge two fascinating fields: machine learning and earth system sciences. My research primarily revolved around tackling the challenges of uncertainty in data, a critical issue in both fields. Here’s an overview of the two key topics I delved into during my PhD.
My first major area of research focused on using semi-supervised learning techniques to enhance plankton classification. In marine science, data labeling is often expensive and time-consuming, leading to datasets that are both sparse and noisy. To address this, I used a technique known as contrastive learning, which leverages unlabeled data to improve the performance of machine learning models. This method helped in pre-training the models using a vast amount of unlabeled data, which was crucial for plankton images where labeled data was scarce. By also incorporating Bayes‘ theorem, we could integrate prior knowledge about species distributions, enhancing the model’s confidence and calibration. This approach significantly reduced the amount of labeled data needed while maintaining high classification accuracy, making it a cost-effective solution for large-scale marine biodiversity studies.
The second focus was on using generative neural networks to quantify and convey uncertainties in environmental data. Generative architectures have traditionally been used to create realistic data samples, but their potential to represent uncertainty has been underexplored. I developed a novel ensemble method called MARVEL, which quantifies uncertainty in predictions. This method proved more computationally efficient than traditional numerical ensemble methods and provided nuanced, calibrated ensemble predictions essential for fields like climate science and meteorology. By accurately representing the uncertainties, these models help in making more reliable predictions, crucial for decision-making in environmental science.
My PhD journey was a deep dive into addressing the complexities of uncertainty in machine learning models, with applications that hold significant promise for marine biology and environmental sciences. My research offers new pathways for scientific inquiry and practical applications, providing hope in fields where data uncertainty is a significant challenge.
The techniques developed during this research hold promise for a wide range of applications, from improving biodiversity assessments in marine environments to enhancing predictive models in climate science.
