Supervised, unsupervised, and a peek at reinforcement

Three families of machine learning, told apart by one question - do your examples come with an answer, with nothing, or with a reward?

Every machine-learning method you will ever meet belongs to one of three families, and you can sort them with a single question: when the model looks at an example, does it come with the right answer attached, with nothing attached, or with a reward that scores how good its last move was? That one difference - labels, no labels, or a reward signal - decides almost everything about how the thing learns.

One question splits the whole field

The names sound intimidating, but the dividing line is mundane. Supervised learning is studying with an answer key: every example is a pair, an input and the correct output, and the model's whole job is to copy the mapping from one to the other. Unsupervised learning is handed a pile of examples with no answers at all and asked to find the structure hiding inside - which things clump together, which directions matter. Reinforcement learning gets neither; it acts, receives a number saying how well that went, and adjusts to earn more of it.

See it: same points, two tasks

The fastest way to feel the split is to hand one pile of points two different jobs. Below are 44 points that naturally form two blobs. Drag the slider from 0% to 100% to control how many of them come with a label. At 0% the machine has no answers, so it does the only thing it can - it invents the two groups by proximity (that is k-means clustering). At 100% every point wears its true colour, so the task flips: the machine now just learns the boundary between the two known classes. Same points; the labels are what change the job.

Notice what the two extremes are really doing. At 0% the machine never sees the truth, so k-means can misplace points in the overlap - it is guessing the groups, and a couple of guesses may disagree with the real colours. At 100% there is no guessing left: every label is given, and the model only has to draw the line where a point is equally close to the two class means ${\mu }_A$ and ${\mu }_B$ - the set of points with $\Vert x-{\mu }_A\Vert =\Vert x-{\mu }_B\Vert$. In between you are in semi-supervised territory: a few labels sharpen a boundary while the raw structure is still doing some of the work.

The distinction, in one equation each

The cleanest way to see that these are genuinely different problems is to write down what each one is trying to minimise or maximise. Watch where the label $y$ appears - and where it vanishes.

A peek at reinforcement

The scatter above can't show the third family, because reinforcement learning has no fixed dataset to slide through. Instead there is a loop: the model (the policy) sees a state, picks an action, and the world hands back a new state and a reward. No single step is ever labeled "correct" - the model must figure out, from delayed and noisy scores, which choices paid off. It is how systems learn to play Go, steer robots, and - in Lesson 53 - how RLHF nudges a language model toward answers humans prefer. Notice how the code stops being a call to fit(X, y) and becomes a feedback loop.

# supervised: X has features, y has the answers
model.fit(X, y)                 # learn the mapping from the labels

# unsupervised: only X, no answer column exists
kmeans.fit(X)                    # the machine invents the groups

# reinforcement: no dataset at all - a reward loop
for step in range(10000):
    action = policy(state)          # act on what you see
    state, reward = env.step(action)  # feedback, not a label
    policy.update(reward)          # chase more reward

Check yourself