What it is
Unsupervised learning finds structure in data without labels. Transfer learning reuses what a model learned on one big task as the starting point for another. Together they are the idea behind pre-training: learn from huge amounts of unlabelled data first, then adapt.
What it is for
When you have few labelled examples, start from a model pre-trained on lots of data (a language model, an image model) and adapt it. This is how most practical AI is built today.
How to do it
- Find a pre-trained model for your kind of data (text, images, audio) on a model hub.
- Decide how much to change: use it as it is, train only a small new part on top, or fine-tune all of it.
- Prepare your small labelled dataset.
- Train with a small learning rate, so you adapt the model without erasing what it already knows.
- Evaluate on data from your real use case: if it is very different from the pre-training data, expect to need more care.