What Is Machine Learning?

Programming, inverted

Traditional programming: a human writes rules, the computer applies them to data to produce answers.

Rules + Data → Answers        (classic programming)
Data + Answers → Rules        (machine learning)

Machine learning flips it: you show the computer many examples of inputs with the correct answers, and it works out the rules itself. The learned rules are called a model.

Try writing rules to recognize a cat in a photo: "has pointy ears"? (so do foxes) "has fur"? (not in silhouette)… Rule-writing collapses immediately. But show a learning algorithm a million labeled photos and it discovers, on its own, the statistical patterns that separate cats from everything else. ML shines exactly where rules are hard to write but examples are easy to collect.

The formal definition

Tom Mitchell's classic definition (1997): a program learns from experience E with respect to task T and performance measure P, if its performance at T, measured by P, improves with E.

Spam filtering: T = classify emails, P = % classified correctly, E = a pile of emails hand-labeled spam/not-spam. Every ML project decomposes into these three questions — what's the task, how do we measure success, what experience (data) do we have?

When to use ML (and when not to)

Use ML when:

Rules are too complex to write (vision, speech, language)
The problem changes over time (fraud patterns, recommendations)
Personalization at scale is needed (each user effectively needs their own rules)

Do NOT use ML when:

Simple rules already work ("flag orders over $10,000" needs no neural network)
You have almost no data
Every decision must be exactly explainable and auditable by law
Errors are unacceptable and there's no human fallback

A professional's first instinct is to ask "do we even need ML here?" — often the honest answer is no.

The vocabulary you'll use forever

Term	Meaning	Example (house prices)
Example / sample	one data point	one house
Features ( $\mathbf{x}$ )	the input measurements	size, age, location
Label ( $y$ )	the answer to predict	sale price
Model ( $f$ )	the learned function	$\hat{y} = f(\mathbf{x})$
Training	fitting the model to data	gradient descent
Inference	using the trained model	pricing a new listing

Key takeaways

ML learns rules from examples instead of having them programmed.
Define task, performance measure and experience before touching any algorithm.
ML is a tool, not a religion — simple rules beat models more often than beginners expect.