Machines learn in different ways, and the right method depends on the data available and the question asked. Every approach can be grouped into three families: supervised, unsupervised and reinforcement learning.
Where they differ is the fuel they use. Supervised learning trains on labelled examples, where each case already has its answer. Unsupervised learning looks for structure in unlabelled data. Reinforcement learning learns through trial and error, guided by rewards from a surrounding environment.
Supervised learning — learning with answers
- Classification: assigning each case to a category, such as benign or malignant
- Regression: predicting a continuous value, such as a tumour size
- Remarkably effective for medical diagnosis and image classification
Unsupervised learning — finding structure alone
- Clustering: grouping similar cases, such as patient subgroups
- Anomaly detection: flagging cases that stand out from the norm
- Useful for discovery when no labels exist
Reinforcement learning — learning from feedback
- An agent acts in an environment and is rewarded for good actions
- Excels at sequential decisions, such as treatment planning
- Less common in imaging today, promising for therapy optimisation