Foundation Model
Also called: base model, pretrained model
A large model pretrained broadly on general data, then adapted to specific downstream tasks rather than trained for one task from scratch.
A foundation model is trained once on a very large general corpus, then adapted through fine-tuning or prompting to many specific tasks. Medical foundation models exist for text, imaging, and pathology.
The appeal is data efficiency. Building a task-specific model requires labeled clinical data, which is expensive and often unobtainable in rare conditions. Adapting a pretrained model can reach useful performance with far fewer labels.
Where This Gets Misread
Regulation assumes a fixed device with a defined intended use, and a foundation model adapted to twenty downstream tasks does not fit that shape. Every adaptation also inherits whatever the pretraining corpus contained, including biases nobody in the adapting organization can audit.