Types of learning

Five categories, one underlying question

Every category of learning answers the same underlying question differently: what kind of feedback does the model get while it learns? That single axis is what separates all five categories covered here, and keeping it in view makes the taxonomy far easier to hold onto than memorising five separate definitions.

Supervised learning: a label for every example

Supervised learning gets the most direct feedback possible: for every training input, the correct output is given. Predicting a house’s sale price from its features, or classifying an email as spam or not, are both supervised problems — the training data already contains the answer the model is meant to produce. Nearly every lesson in the supervised models group of this roadmap sits squarely in this category, and it is where most classical machine learning work actually happens, because labelled data, where it exists, makes the learning problem by far the most tractable of the five.

Unsupervised learning: no labels at all

Unsupervised learning gets no answer key whatsoever — only raw inputs — and has to find structure on its own: which points naturally group together, or which few dimensions capture most of what varies across a dataset. There is no “correct” clustering to check against, which is exactly why the unsupervised learning group later in this roadmap spends real time on what it even means to evaluate a result when no ground truth exists.

Semi-supervised learning: the realistic middle

Most real projects do not land cleanly at either extreme. Labelling data costs time and often needs a domain expert, so a project frequently ends up with a small labelled set and a much larger pool of raw, unlabelled data sitting unused. Semi-supervised learning uses both: the labelled portion anchors the model to genuinely correct answers, while the much larger unlabelled portion still contributes information about the overall shape of the data, often substantially improving on what the small labelled set could achieve alone.

Self-supervised learning: inventing your own labels

Self-supervised learning solves the labelling problem by sidestepping it: rather than a human assigning labels, the task is constructed directly from the data’s own structure. Hide a word in a sentence and ask a model to predict it from the surrounding words — the sentence itself supplies the correct answer, with no human labelling involved at all. This is, concretely, how the language models behind Upshard’s separate AI Engineer roadmap are pretrained at the scale they operate at: there is nowhere near enough human-labelled text in existence to train them the supervised way, so the vast bulk of their pretraining is this self-supervised prediction task, repeated across an enormous volume of ordinary text.

Reinforcement learning: feedback from consequences

Reinforcement learning is the odd one out structurally: there is no dataset of correct answers prepared in advance at all. Instead, an agent takes actions inside an environment and receives a reward signal afterward, learning over many attempts which actions tend to lead to good outcomes. A program learning to play a game by trial and error, improving as it accumulates experience of what wins and what loses, is the clearest everyday example — covered in more depth in this roadmap’s own reinforcement learning lesson.

What to watch for

The video works through three of these five categories — supervised, unsupervised and reinforcement learning — each with an actual worked example rather than a bare definition, which is the right level of depth for a first pass. No single video found anywhere covers all five categories properly in one place; semi-supervised and self-supervised learning, covered above, are the two this lesson’s written material carries that the video does not.