Transfer learning uses features learned on large datasets (ImageNet) for new tasks.
Feature extraction: Freeze pretrained layers. Train only new classification head. Fast, works with little data.
Fine-tuning: Unfreeze some layers. Train end-to-end with low learning rate. Better performance, needs more data.
When to fine-tune: When your domain differs from ImageNet. Medical images, satellite imagery.
Interview tip: Start with frozen backbone. If results are poor, gradually unfreeze layers from top to bottom.