Definition
Transfer learning uses representations, parameters, or other learned structure from a source task or domain to help a target task or domain. Instead of learning the target model entirely from its target data, the procedure starts from something learned elsewhere.
The transferred model can remain frozen as a feature extractor, be fine-tuned in full, or be adapted through a small set of new parameters such as LoRA matrices. Transfer may cross tasks, data distributions, or both. Pretraining followed by downstream fine-tuning is a common modern pattern, but it is one case within the broader category.
Sinno Jialin Pan and Qiang Yang's 2009 survey organized transfer learning by how source and target domains and tasks differ and by whether labeled target data is available. The survey is a vocabulary and literature synthesis, not the origin of transferring learned structure between problems.
Transfer helps when the source has reusable structure. It can hurt when source features, labels, or shortcuts do not match the target. This negative transfer may appear as lower target accuracy, slower convergence, or brittle behavior on a subgroup. Evaluation must compare against a target-only baseline and cover the deployment distribution.
Distinguish it from nearby terms
Fine-tuning continues parameter training and often performs transfer. Domain adaptation focuses on distribution shift between source and target. In-context examples can change behavior without transferring learned parameters because the model weights stay fixed.
Check your understanding
A vision encoder pretrained on everyday photographs is reused for medical images. Which source assumptions might fail, and what comparison would reveal negative transfer rather than ordinary target-data scarcity?