Foundation Models for Human Sensing with Support for Fine-Tuning
Opis
This service enables the use of photoplethysmogram (PPG) foundation models for various tasks. The user provides a modestly sized labelled dataset for some task of their interest (e.g., blood pressure estimation) on PPG data. They can select one of the offered foundation models to serve as the basis for their task-specific model. The foundation model is then fine-tuned on this dataset. The user can also configure the fine-tuning parameters; defaults are provided. The expected performance exceeds the performance of a model solely trained on the user's dataset, especially in terms of generalization and robustness. Example datasets and fine-tuned models are provided. Three models for PPG data are offered initially: PaPagGei, GPT-PPG, and Pulse-PPG. The first two are more appropriate for clinical data, and the third for data from wearables. This selection may be extended later, including to models for other types of sensors. Fine-tuning parameters include different possible heads, different amounts of unfreezing (head only vs. head and layers of the foundation model), and are based on the PyTorch framework.
Pričakovani rezultati
A fine-tuned model adapted to the user's task, based on the robust and general representations obtained through existing pre-trained foundation models.
Kartica storitve
| Kategorija | AI development and deployment |
| Domena | Sensing, Health Monitoring |
| Vertikala | Health and Biotechnology |
| Ciljni uporabniki | Manufacturers and Researchers of Health Monitoring Devices and Software |
| Tip storitve | Models, Fine-tuning Workflow |
| Rezultat | Fine-tuned model and visualisation of which parts of the signal are important for the prediction. |
| Zahteve | Requires NVIDIA GPU node. Sufficient amount of sensor data with labels for the task of interest in the prescribed format is required for fine-tuning. |
| Omejitve | |
| Na voljo od | Q4/2026 |
| Licenca | BSD, MIT, GPL |
| Slika | assets/service-visuals/access-to-foundation-models-for-human-sensing-with-support-for-fine-tuning.png |
| Oznaka slike |