Machine learning, explained

Lesson 1 of 81956–19636 min

What machine learning is

On 24 February 1956, a checkers program running on an IBM computer was shown on television. It was the work of Arthur Samuel, an engineer who had joined IBM in 1949. The rules of checkers were the easy part; any programmer could type them in. What made the program remarkable was that it got better by playing.

In July 1959 Samuel described the work in the IBM Journal under a title that named a field: “Some Studies in Machine Learning Using the Game of Checkers”. He reported that 8 to 10 hours of machine playing time were enough for the program to play a better game than the person who wrote it.

In 1962 it beat Robert Nealey, billed as a checkers master. Later analysis showed Nealey was not quite the master the publicity claimed, and in a 1963 rematch he won one game and drew five. The idea outlived the hype.

Rules versus learning

Ordinary software is rules-based: a programmer works out the logic and writes it down. if amount > limit then reject. That works when the rule is known, stable and short. It breaks down when nobody can write the rule: what makes a checkers position strong, an email spam, or a transaction suspicious? Machine learning turns the problem around. You supply examples and a measure of how wrong the program is, and an algorithm adjusts the program until the error is small. The result is a model: a function whose behaviour was fitted from data rather than typed in.

What a model actually is

Samuel's program scored board positions with a formula: a weighted sum of features such as piece advantage and mobility. He described two kinds of learning. Rote learning stored positions it had already evaluated so it could look them up later. Generalisation learning adjusted the weights of the scoring formula, and swapped terms in and out, according to how games turned out. That second idea is the core of nearly every model in use today: a set of features that describe the input, a set of parameters (weights) that combine them, and a training procedure that nudges the parameters to reduce error.

  • Supervised learning: every example comes with the right answer, a label, and the model learns to predict it. Spam filters, price forecasts, fraud scores.
  • Unsupervised learning: no labels. The model finds structure on its own, such as groups of similar customers or readings that look abnormal.
  • Reinforcement learning: an agent learns from rewards by acting, the way a program can improve through thousands of games against itself.

When to use it, and when not

Machine learning earns its cost when the pattern is real but too complex to write by hand, when there is enough representative data to learn it from, and when an occasional wrong answer is acceptable. It is the wrong tool when a rule is mandated (a legal limit, a tax rate), when there is little data, or when every decision must be explained line by line. A model is a statistical bet, right most of the time on data that looks like what it was trained on. Everything in the rest of this path is about making that bet well.

Key terms

Machine learning
Building a program's behaviour by fitting it to examples instead of writing explicit rules.
Model
The fitted function that turns inputs into predictions.
Feature
A measurable property of the input, such as piece advantage or the amount of a payment.
Parameter
A value inside the model, such as a weight, that training adjusts.
Training
Repeatedly adjusting the parameters to reduce the model's error on examples.

Check what you learned

  1. 1.What distinguishes a machine-learning system from rules-based code?

  2. 2.In Samuel's checkers program, what did generalisation learning adjust?

  3. 3.A bank must reject any transfer above a statutory limit. What is the best approach?

Practise it for the exam

Sources