They are two ways to train a machine learning model. In supervised learning, each example comes with the right answer, the label: a photo and the word "cat". In unsupervised learning, the examples have no label and the model looks for groups or patterns on its own, like customers who buy alike.
A question from the app. Pick an answer.
You train a model on thousands of photos labeled "cat" or "dog". What learning is it?
Each example comes with the right answer, the label. The model learns to give that answer for photos it never saw.
There the examples have no label: the model looks for patterns on its own, like grouping similar photos without knowing what they are.
That one learns by trying actions and getting rewards or penalties, like a game. Here it does not act: it looks at photos with their answer.
You give it thousands of emails labeled "spam" or "not spam", and the model learns to label new emails. Or you give it houses with their sale price, and it learns to estimate the price of a house it never saw. In the first case it picks a category (a classifier); in the second, it computes a number.
The expensive part is the labels: often a person adds them, one example at a time. And if the labels are wrong, the model learns the mistake.
A store gives a model its customers' purchases without saying what to look for. The model puts together people who buy alike: some get diapers every week, others only show up for sales. What each group means and what to do with it is up to a person afterwards.
It also helps find the unusual: a card payment that looks like none of the usual ones may be fraud, even if nobody ever labeled that kind of fraud before.
Reinforcement: the model tries actions and gets rewards or penalties, like someone learning a video game. Self-supervised: the labels come from the data itself. That is how text generating LLMs are pretrained: the model tries to predict the next token from the previous ones, and the answer is already in the text, with nobody labeling anything by hand.
One model can go through several: a chat like ChatGPT is first pretrained this way, then tuned with answers written by people (supervised) and with their ratings (reinforcement). All of that is training; when you use it, what you tell it helps while it stays in the context, but it is not written into its weights.
"Without labels it cannot learn." It can: unsupervised learning learns patterns and similarities without anyone marking the right answer first. "Supervised means a person watches it while it learns." It means the examples come with the answer. "Everything is either supervised or unsupervised." There is also reinforcement and self-supervised learning, and they get combined.