Learn

Machine learning, explained

From a 1950s checkers program to models running in production: learning from data, splitting it, scoring it, and keeping models honest once they ship.

8 lessons · 53 min

  1. 1What machine learning isFreeArthur Samuel's checkers program improved by playing, and in 1959 his paper named the field. What learning from data means, and when rules are better.1956–1963 · 6 min
  2. 2Supervised learning: features and labelsIn 2002 Paul Graham let 8,000 labelled emails write his spam filter. How supervised learning maps features to labels, and regression versus classification.2002 · 6 min
  3. 3Unsupervised learning: clusters and anomaliesA Bell Labs draft from 1957 gave us k-means. How models find groups in unlabelled data, and how the same idea spots the transaction that does not fit.1957–1982 · 6 min
  4. 4Training, validation and test setsThe Netflix Prize hid its deciding data from every leaderboard. Why models are judged on data they never saw, and how to spot overfitting and underfitting.2006–2009 · 7 min
  5. 5Evaluating models: beyond accuracyA Cranfield librarian gave us precision and recall. The confusion matrix, a worked example in numbers, and why 95% accuracy can mean a useless model.1957–1972 · 7 min
  6. 6Decision trees, random forests and boostingLeo Breiman judged models by how well they predict. One tree, a forest of them, and boosted trees that fix each other's mistakes, without the maths.1984–2016 · 7 min
  7. 7Data quality, bias and leakageAmazon's hiring model learned to mark down women because its data told it to. How flawed data, inherited bias and leaked answers make models wrong.1990s–2018 · 7 min
  8. 8Models in production: MLOps, drift and retrainingGoogle Flu Trends worked, then quietly drifted until it doubled the real figures. Deploying models, watching them for drift and retraining them safely.2008–2015 · 7 min