Justus’ academic training spans differential geometry, topology, statistics, and machine learning, combining rigorous mathematical reasoning with applied computational methods. Their Master’s thesis focused on the Bochner technique and curvature conditions in differential geometry. Subsequently, they worked as a research engineer in industrial predevelopment, translating state-of-the-art machine learning research into deployable systems. Their work concentrated on computer vision, reinforcement learning, and multimodal sensor data that included optical, ultrasound, and mechanical signals, contributing to applied research projects, proof-of-concept implementations, and collaborations with industry and research partners.
Currently, Justus’ PhD research focuses on continual few-shot learning and biologically inspired neural architectures, with a strong emphasis on computer vision and the adaptation of foundation models. Their work includes Comply, an efficient fruit fly-inspired architecture that outperforms significantly larger transformer models; Robust Weight Imprinting, a framework for adapting frozen foundation models to new classes without gradient training; SCAM, the largest dataset of real-world typographic attacks, exposing vulnerabilities in vision-language models; and SynthMT, synthetic data that lets foundation models segment microtubules at human level.
In the SWAMBIT project, Justus studies programs of layers in large language models: skipping and repeating the layers of a frozen, pretrained foundation model depending on the input, and relating the resulting computation to cortical areas in the brain.
Beyond research, Justus is actively involved in the Berlin AI community through BLISS, where they help organize reading groups, community events, and the BLISS AI Speaker Series, fostering exchange between academic research and applied machine learning practice.