core Estimated learning time: 5 h

6.14 Transfer learning and fine-tuning

You can hit strong accuracy on 500 labelled images.

Before:05. Classical Machine LearningUnlocks:07. Natural Language Processing 11. The ECE Bridge — Signals, Edge & Embedded AI

Transfer learning is the working default of applied deep learning: start from pretrained weights, freeze, fine-tune progressively — strong accuracy on five hundred labelled images is routine this way and impossible from scratch. It sits right after the architectures it reuses. The destructive mistake is fine-tuning everything at a high learning rate immediately, which erases the pretrained features before they can help; discriminative rates and gradual unfreezing exist for exactly that reason.

Work through these

Feature extraction vs. full fine-tuning
Using a pretrained network as a fixed feature extractor, or continuing to train all of it, are two ends of a range with different data requirements. Choosing correctly is the main decision in this topic.
Full Stack Deep Learning · Course
Layer freezing and discriminative learning rates
Holding early layers fixed and training later ones, possibly at different step sizes. This is what makes fine-tuning work on a few hundred examples.
Domain shift and when transfer fails
Transfer helps when the source and target data resemble each other and can hurt when they do not. Knowing the failure case prevents a confident waste of time.
timm and pretrained model zoos
The libraries and public collections that make thousands of pretrained models available. Starting from one of these is the normal practice rather than an advanced technique.

Free resources

Links last checked 29 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.