In transfer learning, a model that has already learnt a large general task serves as the starting point for a new, specialised one. Instead of starting from scratch, it is adapted, for instance through fine-tuning, using comparatively little data of your own. This saves computing time and data, for example when a general image model is meant to learn to spot production defects in photos.
Fine-tuning means further training an already pre-trained AI model on additional, usually task-specific data. Its internal weights are adjusted so that the model handles, for instance, a particular writing style, specialist vocabulary or output format more reliably. Fine-tuning requires carefully prepared example data, computing power and testing. For many applications, well-written instructions and connected knowledge are sufficient and easier to change.
A foundation model is a large AI model pre-trained on very extensive, wide-ranging data that can then be used for many different tasks. The term was coined at Stanford University in 2021. Well-known examples include large language models such as Claude or GPT, as well as image models. Businesses usually build their applications on such models rather than training their own from scratch; the EU AI Act refers to general-purpose AI models.
Parameter-efficient fine-tuning (PEFT) covers techniques that adapt a large pre-trained model to a task by training only a small portion of its parameters or adding a few extra ones. Well-known methods include LoRA, QLoRA, adapters and prefix tuning. This greatly reduces computing and memory requirements, and small add-on modules can be stored for several tasks instead of full model copies. Example: a language model is adapted to the style of internal technical documents.
