Artificial intelligence (AI) concepts are propelling nearly all computer vision-intensive applications in life science, biomedical research, space exploration, high-tech manufacturing, and security technology. While traditional image processing methods are based on linear space-invariant assumptions, neural networks are inherently non-linear and have the potential to outperform these methods. Neural networks are trained to perform a certain task using very large sets of data. The feature of adapting to data by extracting the essential information and using it to form decisions or make predictions in a “black box” is what makes this approach so useful for many applications. For scientific applications, however, this black box causes a serious dilemma: what is gained in performance is lost in interpretability of the solution. Also lost is the ability to integrate existing physical knowledge of the system. The aim of the IRIS lab is to open the black box of AI and develop methodologies for context-independent, knowledge-based learning of imaging systems that will address fundamental challenges in all quantitative imaging applications. The proposed AI-technology will be applied to electron, optical, and ultrasound imaging to unravel dynamic molecular processes in living organisms: conformational ensembles of proteins, single-molecule dynamics in thick tissue and super-resolved vasculature mapping in real-time.