What Is Machine Learning?

Why Does This Exist?

Forty thousand boxes a night move past a camera in a parcel depot, and somebody wants the crushed ones pulled off before a customer opens one. Try writing that as a rule. Outline deviation. Missing corner shadow. Torn tape. You will add condition after condition and still miss the box that got sat on, because crushed was never a threshold on anything you can name.

So stop writing the rule. Take four thousand photos a human already marked crushed or fine, hand the pile to a procedure whose job is finding a function that reproduces those verdicts, and use what comes back.

That function is $f$, from an input $x$, one photo, to an output $y$, crushed or fine. You never see it. You see examples of what it does, and you search a family of candidates for whichever matches best. To generalise is to be right about a photo nobody labelled, which is the only thing anybody pays for. When that goes wrong on new data it has a name, overfitting, and it is the natural next page after this one.

Think of It Like This

How It Actually Works

All four paradigms feed one pipeline: examples in, a fitted function out, that function applied to inputs nobody has seen. What changes is the supervision arriving alongside the input, and that single difference carves the field into paradigms.

The four, on the belt

Supervised. Every $x$ has its $y$. Four thousand labelled photos, a fitted detector, eight hundred held back so you can count how often it is right. Almost everything in production is supervised, and labels are why. They are the expensive ingredient.

Unsupervised. No $y$ anywhere. Push two million unlabelled photos in and ask what structure they have. k-means might return clusters that turn out to be night-shift lighting, the third camera's dirty lens, and genuine damage. Two of those are operations problems, and nothing here has an answer key.

Semi-supervised. Four thousand labelled, two million not. That is what a real project has on day one. The unlabelled mountain still carries information, in where photos bunch together and which directions barely vary, and using it places a boundary with far fewer labels.

Reinforcement. No answer key. A consequence. A robot arm lifting a flagged box earns a reward when it lands in the right chute and nothing when it fumbles. Nobody supplied joint angles, so credit has to spread backward across the whole sequence of moves, which is why reinforcement learning is harder than the tidy diagram suggests.

Where the split stops holding

Blank out a patch of each unlabelled photo and train a model to reconstruct what you removed. There is a label, the patch, and no human produced it. That is self-supervised learning: supervised training on labels invented from the data's own structure. It fits none of the four boxes, and it is how nearly every large modern model gets pre-trained. So treat the paradigms as a question. What supervision do I actually have?

When to write the rule instead

If the rule is writable, write the rule. Parcels over thirty kilograms go to the pallet lane: the belt has a scale, the threshold is a policy decision, and a lookup wins on every axis. Machine learning earns its place when the rule exists but nobody can state it: faces, speech, whether a box looks crushed. Elsewhere you bought a training pipeline and an unexplainable failure mode in exchange for reproducing an if statement.

Code

One measurement per parcel, one threshold, learned rather than guessed.

1import numpy as np2
3
4def best_cut(x: np.ndarray, y: np.ndarray) -> float:5    cuts = np.unique(x)6    # One matrix compare stands in for the whole search over candidate rules.7    wrong = ((x[None, :] >= cuts[:, None]) != y[None, :]).sum(axis=1)8    return float(cuts[wrong.argmin()])9
10
11rng = np.random.default_rng(0)12dent = np.concatenate([rng.normal(2.0, 1.0, 400), rng.normal(6.0, 1.0, 400)])13damaged = np.concatenate([np.zeros(400), np.ones(400)]).astype(bool)14
15for name, cut in (("hand-written rule", 3.0), ("learned rule", best_cut(dent, damaged))):16    print(f"{name:18s} cut at {cut:5.2f}   {((dent >= cut) == damaged).mean():.3f} correct")17    # -> hand-written rule  cut at  3.00   0.921 correct18    # -> learned rule       cut at  4.06   0.984 correct

Eyeballing the histogram gets you 3.0 and 92.1 percent. The search gets 4.06 and 98.4 percent, about two and a half thousand fewer mistakes a night. No neural network anywhere: it defined a family of rules, scored every member, kept the winner. Everything else in the field is a bigger family, and the simplest fitted rule of that kind is linear regression.

Watch Out For

The Quick Version

• Machine learning fits a function from input to output using examples, instead of you writing the mapping.
• The paradigms differ by one thing: what supervision arrives with the input. A label for every example, a few, none, or a reward.
• Semi-supervised is the realistic starting condition for most projects, not the exotic case.
• Self-supervised learning invents labels from unlabelled data, so the classic three-way split no longer covers the field.
• Unsupervised learning has no answer key, which makes evaluation a judgement about usefulness.
• If the rule is writable, write the rule. Everything here also assumes training and deployment data match, and that assumption fails first.

What to Read Next

• Overfitting and Underfitting
• Linear Regression