Machine Learning is the science (and art) of programming computers so they can learn from data. Pioneer Arthur Samuel put it simply in 1959:

“[Machine Learning is the] field of study that gives computers the ability to learn without being explicitly programmed.”

Tom Mitchell gave a more engineering-flavored definition in 1997:

“A computer program is said to learn from experience E with respect to some task T and some performance measure P, if its performance on T, as measured by P, improves with experience E.”

Your spam filter is a good example. Given examples of spam emails (flagged by users) and regular (“ham”) emails, it learns to flag new spam. Here:

• Task (T) — flag spam for new emails
• Experience (E) — the training data (past emails, labeled spam/ham)
• Performance (P) — accuracy: the ratio of emails classified correctly

  flowchart LR
  T["Task (T)
flag spam"] --> P
  E["Experience (E)
labeled emails"] --> P["Performance (P)
accuracy improves"]

If you just download a copy of Wikipedia, your computer suddenly has a lot more data — but it isn’t any better at a task. That’s why downloading data is not Machine Learning; learning means improving at T, measured by P, as you get more E.

The idea in one sentence

Machine Learning is a way to build programs that learn a mapping from inputs to outputs from examples, instead of hard-coded rules.

In ML, we usually have:

• Inputs (features): the data we observe (like age, clicks, temperature)
• Outputs (labels/targets): the thing we want to predict (like price, spam/not spam)
• A model: a mathematical function with adjustable parameters
• Training: the process of adjusting parameters so predictions match reality

A good mental model

Think of ML like learning a recipe from tasting many dishes.

• Traditional programming: you write the recipe.
• ML: you taste many examples (data), and the model approximates the recipe.

What problems is ML good for?

ML shines when:

• the rules are too complex to write explicitly
• the patterns are statistical and noisy
• you want the system to improve with more data

Examples:

• email spam detection
• product recommendations
• speech-to-text
• fraud detection
• predicting house prices

To summarize, from the book, Machine Learning is great for:

• Problems needing lots of fine-tuning or long lists of rules — one ML algorithm can often simplify code and perform better than the traditional approach (spam filters are the classic example).
• Complex problems with no known algorithm — e.g. speech recognition. You can’t hand-write rules for “high-pitch sound = the word two” that scale to millions of voices and languages, but a model can learn it from recordings.
• Fluctuating environments — an ML system can retrain and adapt to new data, instead of being manually patched forever.
• Getting insights about complex problems and large amounts of data — inspecting what a trained model learned can reveal patterns a human would never spot. This is called data mining.

ML is not magic

ML does not automatically mean:

• correct answers
• unbiased answers
• causation (often it learns correlations)

You still must:

• choose good data
• evaluate correctly
• monitor for drift and failures

Core vocabulary (keep these handy)

• Feature: an input variable.
• Target / label: output variable.
• Dataset: many examples (rows) with features (columns).
• Training: fitting a model.
• Inference: using the trained model to predict.

Tiny checkpoint

Can you explain ML like this to a friend?

“ML is when we show a program many examples so it can learn patterns and make predictions on new data.”

🧪 Try It Yourself

Exercise 1 – Train-Test Split

Exercise 2 – Fit a Linear Model

Exercise 3 – Evaluate with MSE