How AI learns
AI can learn from examples, from feedback or by following rules.
The chapter at a glance
Examples or experience
How does it learn?
With labelsSupervised learningWithout labelsFinding patternsWith rewardsReinforcement learningWith rulesSymbolic AI
Prediction or decision
Explanation and checking
In this chapter
- 1What is machine learning?The computer learns patterns from examples instead of a programmer writing every rule.
- 2Three ways to learn: with labels, without them, and from the data itselfDifferent methods use different kinds of feedback.
- 3What is deep learning?Deep learning is a sub-field of machine learning that uses neural networks with many layers.
- 4How does a neural network work?A neural network turns input into output through layers of connected computing units.
- 5What is reinforcement learning?The system chooses actions and gets feedback on how well they worked.
- 6Symbolic AI: rules, logic and knowledgeAI is not only about huge networks. There are also systems that work with clear rules.
- 7Search, planning and optimisationThese classic AI sub-fields answer the question: which option should we choose, and in what order?
- 8Causality: what causes what?Most patterns in data do not automatically tell us what a change will cause.
- 9Bayes and probability: how to reason under uncertaintyMany AI systems have to estimate what is likely when we do not know everything.
- 10Decision trees and classic ML modelsPlenty of tasks do not need huge neural networks.
- 11Knowledge graphs: how are facts connected?A knowledge graph describes things and the connections between them.
- 12AI for networks: graph machine learningGraph learning uses patterns between connected things, not just within individual rows of data.
- 13Evolutionary computing: try, select, improveSome algorithms search for good solutions by imitating the processes of evolution.
- 14Classification, recommendations and unusual eventsA large part of AI does not create content; it helps recognise or sort cases.
- 15Forecasting over time: from sales to energyA time series is a set of measurements recorded in sequence over time.
- 16Can a model learn without collecting all the data in one place?Federated learning lets several devices or organisations train a model together without necessarily pooling all their raw data.
- 17What is explainable AI (XAI)?Explainability helps us understand why a system suggested a particular prediction or decision.