A pretrained network with every block frozen (locked) except the new head at the end. Training just the head is fast and needs very little data.
Freeze the entire backbone and train only the new head. Fast, needs very little data, and impossible to break the pretrained features.
