Degrees

Machine learning

Loss functions, gradient descent, logistic regression, regularisation, trees, clustering, backpropagation and cross-validation, with the maths worked out.

Course · 8 lessons · 61 min

Taught inData ScienceArtificial Intelligence

  1. 1The learning problem and loss functionsFreeData drawn from an unknown distribution, a family of hypotheses and a loss: the formal setup of supervised learning, and why the choice of loss matters.1984–today · 7 min
  2. 2Linear regression and gradient descentFit a line by minimising squared error, solve it in closed form, then solve it again by walking downhill: the gradient descent step worked by hand.1805–1847 · 8 min
  3. 3Logistic regression and classification metricsTurn a linear score into a probability with the sigmoid, train it with log loss, and judge it with precision, recall, F1 and ROC instead of accuracy alone.1958–today · 8 min
  4. 4The bias–variance trade-off and regularisationWhy a model can fail by being too simple or too flexible, how expected error splits into bias, variance and noise, and how ridge and lasso tame it.1970–1996 · 8 min
  5. 5Decision trees and ensemblesSplit the data greedily on the most informative question, measure it with entropy, then combine many trees by bagging or boosting.1986–2001 · 8 min
  6. 6k-nearest neighbours and k-meansTwo algorithms built on distance: k-NN classifies a point by its neighbours, k-means groups unlabelled points around centroids. Both live or die by scaling.1967–2007 · 7 min
  7. 7Neural networks and backpropagationLayers of weighted sums and non-linearities, trained by the chain rule run backwards: the forward pass, the backward pass and one neuron updated by hand.1970–1986 · 8 min
  8. 8Evaluation and cross-validationHow a model will do on unseen data: holdout, k-fold, stratified and time-series splits, nested cross-validation and how leakage fakes a good score.1974–today · 7 min