Appl-FM — Berlin Initiative for Applied Foundation Model Research

Foundation models are large, general-purpose machine learning models. They are trained once on broad data and then adapted to many different tasks. Like electricity or the steam engine, they show the hallmarks of a general-purpose technology: they are improving rapidly, spreading across fields, and spawning complementary innovations. Adapting them to new applications, however, is far from trivial and is sometimes impossible. This blocks their use in most industrial, medical, and scientific settings, and it is exactly the problem Appl-FM addresses.

The initiative works on three research scenarios, each with its own challenge:

• Robotics: teaching robots human manipulation and locomotion skills by observing demonstrations, instead of through time-consuming recordings with human interaction.
• Quantitative biology and smart microscopy: extracting reliable quantitative information from imbalanced, rare imaging data in clinically relevant fields.
• Predictive medicine: supporting doctors with trustworthy decision support across a broad range of diseases, based on few, multimodal data.

Despite their differences, the three scenarios share core requirements. Models must learn from rare, sparse, and continually arriving data, handle multiple modalities, and be robust, fair, and explainable. To meet these requirements, Appl-FM focuses on four methods: brain-inspired architectures for continual few-shot learning, data-centric vision and motion learning with human feedback, automated holistic evaluation of robustness, fairness, and explainability, and the adaptation of text embeddings in multimodal settings.

In the long term, the initiative aims to make foundation models more accessible. Today, a few large technology companies dominate their development because they control the data, compute, and deployment infrastructure. Appl-FM counters this by developing methods to adapt publicly available models, both for well-studied domains and for neglected niche ones. It also strengthens the role of interdisciplinary academic research in a field that has so far been driven almost entirely by industry.

Within Appl-FM, two DATEXIS PhD researchers work across the research scenarios and methods. Jan Frick, supervised by Prof. Dr. Alexander Löser, works on predictive medicine. His focus is diagnosis prediction and medical reasoning from clinical text, including a benchmark that compares generative and encoder-based language models for clinical outcome prediction. His research is now moving toward self-improving systems and memory management at run time: models that keep learning and refining their knowledge while they are deployed, a direct answer to the challenge of continually arriving data.

Justus Westerhoff, supervised by Prof. Dr. Felix Gers, focuses on biologically inspired machine learning and computer vision. His work covers brain-inspired representation learning, such as sentence embeddings modeled on fruit fly olfaction, and robust few-shot adaptation of foundation models. It also includes real-world robustness evaluation of multimodal models and, in quantitative biology, foundation-model-based microscopy analysis trained on synthetic data. Together, these contributions span few-shot and continual learning, vision, holistic evaluation, and text embeddings, the core methods of the initiative.