Machine learning has enabled capabilities of visual understanding that have been illusive for decades. Similarly, machine learning has enabled new visual synthesis and simulation algorithms that show unprecedented realism and computational efficiency. However, scaling these initial case-specific successes up to complex entire real world scenes remains a major open challenge. Further, current AIML methods are neither robust nor do they lend themselves for the inclusion of prior knowledge and interpretability. Additionally, research related to privacy and security concerns such as model stealing or the inference of privacy relevant information are in their infancy. This RA aims to address both, enhanced functionally as well as trustworthiness, of machine learning driven visual understanding and synthesis.

As machine learning models trained on big data are being used to assist or replace human decision makers in domains that affect people’s lives, concerns are being raised about the ethics of such data-driven decision making. Similar concerns are fueled by the possibilities offered by advanced learning-based methods for understanding from and synthesis of images and videos. Perhaps, not surprisingly, these developments have been characterized by an increasing number of missteps, from discriminating minorities, fueling the spread of misinformation and increasing societal polarization. This RA will focus on human-centered machine learning models and algorithms, specifically designed to avoid the above missteps moving forward and guarantee fairness, accountability, transparency, explainability, and privacy for all.