Glossary · Phase 4 · Deep learning
Negative transfer
When features reused from another task make results worse than training from scratch, usually because the old task was too narrow or too different.
Try it now
Two questions decide it: how many labels you have, and how close the pretrained task is to yours.
Learn it properly
Transfer learning · Reuse a trained network: freeze the backbone, train a new head, fine-tune with two learning rates. About 40 minutes.
Other terms from the same lesson
- backbone: The trained layers of a network that turn raw input into features; in transfer learning they're kept, frozen or fine-tuned, under a new head.
- parameter group: A set of a model's parameters that an optimizer updates with its own settings, such as a smaller learning rate for a pretrained backbone.
- transfer learning: Reusing a network trained on one task for another: keep its trained layers, replace the last one, and train with the new task's labels.
Never trained an AI?
Teach a small one to write, right in your browser. It takes about 20 seconds and needs no account.