What machine learning is

Don't write the rules. Write down what a good rule looks like, and let the computer find the one that fits your data.

The whole course, in one idea.

Programming vs. learning

Programming: you know the rule, you write it down.
Machine learning: you don't know the rule, but you have examples of it — and you'd like the computer to recover it.
Worth it when the rule is real but too big or too fuzzy to hand-code (spam, tumors), and it drifts.
Not worth it when the rule is short and stable. Then just write the rule.

History

1959 — Arthur Samuel's self-improving checkers program coins "machine learning."
1958 — Rosenblatt's perceptron learns a boundary from examples, weights nudged by mistakes.
1970s–80s — expert systems try to hand-code knowledge as thousands of if-then rules. They rot.
The comeback: stop writing rules, show the machine examples, let it infer. That bet is the whole subject.

Two flavors, same points

Left: you gave the labels, the model learns a boundary. Right: no labels, the model discovers the groups.

The whole field, one line

\theta^{*} = \arg\min_{\theta} \; \frac{1}{N} \sum_{i=1}^{N} \ell\bigl(f_{\theta}(x_i),\, y_i\bigr)

$f_{\theta}$ — the model. Change it → threshold, tree, network.
$\ell$ — the loss. Change it → what "wrong" means.
$\arg\min$ — the training algorithm. Change it → brute force, gradients.

Pick those three and you've specified a learning method completely.

Learning = reducing loss

Slide the threshold, watch the misclassification rate fall to its minimum. That descent is the entire algorithm.

Fit, predict, evaluate

Both learners score 1.000 test accuracy (fraction of held-out points correct, unitless 0–1) — the toy set is separable by design.
Unsupervised clustering scores adjusted Rand index 1.000 vs the sealed blob ids — structure recovered with no labels.
scikit-learn does both in one line each, through the .fit / .predict interface every later chapter reuses.

Takeaways

Every chapter is the same loop: pick a model family, define a loss, minimize it on data, check it generalizes.
ML is worth it when the rule is too many or too fuzzy to hand-code — and it drifts. Otherwise write the rule.
Data quality beats algorithm cleverness. Almost always.
Learn the .fit / .predict interface once; the whole library is free after that.

1 / 8