For machine learning applications, robustness plays an important role because data in a real-world environment changes frequently and is affected by various sources of interference. At the same time, different data sources are available for many machine learning tasks. Combining these sources increases the available information and therefore generally leads to improved predictive performance as well. To this end, understanding and explaining data fusion processes on the one hand, and using uncertainty quantification in a multimodal setting on the other, provides the opportunity to develop more robust approaches based on multiple modalities. In particular, aspects of robustness that are not present in single modality approaches, such as conflicting data and environmental changes that affect only single modalities, are of particular interest here.
Project runtime: 10/2019 - 10/2022
Partner: TUM - Data Science in Earth Observation
Spokeperson: Jakob Gawlikowski