01
Learning from data
Machine learning changes the design problem. Instead of explicitly programming every rule required for a task, the engineer defines a model family, an objective, an optimisation procedure and a source of experience or data.
The central challenge is generalisation: a model should perform well on data it did not see during training, not merely memorise its training examples.
02
Learning paradigms
- Supervised learning learns mappings from labelled examples.
- Unsupervised learning seeks structure without target labels.
- Self-supervised learning constructs learning signals from the data itself.
- Reinforcement learning learns behaviour through interaction and reward.
03
Objectives and optimisation
θ* = argminθ (1/n) Σᵢ L(fθ(xᵢ), yᵢ) + λΩ(θ)
The loss measures how poorly the model behaves on examples. Regularisation can penalise undesirable complexity. Training becomes a numerical optimisation problem over a parameterised function.
04
Representation learning
A powerful model does not merely memorise examples; it learns internal representations that make useful distinctions easier to compute. Deep networks create multiple layers of learned representations, which is one reason they scale effectively across many tasks.
05
Reinforcement learning
Reinforcement learning treats intelligence as sequential decision-making. The agent repeatedly observes a state, chooses an action and receives an outcome or reward.
V*(s) = maxₐ Σₛ′ P(s′|s,a)[R(s,a,s′) + γV*(s′)]