▶Source model
House-price data identify a reliable slope from several observations.
Transfer Learning with Limited Data
Reuse what generalizes and tune what changes
When target data are scarce, a model trained on a related source task can provide a strong starting point and reduce the number of free parameters.
watchone idea
→manipulateone example
→leave withone intuition
The essential idea: retain transferable structure and fine-tune only the task-specific part.
One Concept · One Example
What to notice
- Source learning uses the larger dataset.
- Related tasks can share useful parameters.
- Fine-tuning limits overfit on the small target set.
A prior learned from another task
Target fine-tuning
One office-rent point adjusts only the bias while retaining that slope.
One target point: reuse or relearn?
Move the single office-rent observation. Compare transfer learning with a line fitted from scratch to that point alone.
Transfer learning treats a pretrained model as informative prior structure and fine-tunes only what the target task needs.