Course provider: Jožef Stefan Institute (JSI)
Instructors: Gašper Slapničar (JSI), Mitja Luštrek (JSI)
Learning objectives: Gain practical knowledge on what FMs exist and where they are useful.
Course content: The course will overview the main methodological ideas on how to build, adapt, and evaluate large self-supervised models for biosignals (e.g., PPG, ECG, EEG …) and motion sensors (accelerometers, gyroscopes) under the messy constraints of real-world wearables (noise, motion artifacts, missingness, device/site shifts, and limited labels).
The training covers modern foundation-model backbones for time series (CNN and ResNet encoders, Transformers, etc.), pretraining paradigms (masked modeling, contrastive/relative contrastive learning), and representation design choices that matter specifically for physiology (beat-synchronous views, morphology-aware learning). Moreover we will highlight state-of-the-art approaches on merging LLMs and sensor-based models that result in hybrid approaches allowing for human-language interpretation of wearable data (e.g., SensorLM).
Learning outcomes: Practical knowledge on what FMs exist, where they are useful, and how to take them beyond default setups to be useful in their domains and problems.