The A-Z of AI

Learning

The different methods used to teach AI.

Two of the most common techniques that AI design teams use to train machine learning systems are supervised and unsupervised learning.

In supervised learning, a system is given reference data, which it can use to look for similar patterns in new data. It learns through a process of iteration: trial and error.

Two brackets beside each other detail the nature of supervised vs. unsupervised learning. Supervised learning contains a single pineapple; unsupervised learning contains multiple fruits of different, weird and wonderful shapes and sizes.

Imagine someone wanted to teach a system to recognize different fruits. You might start by showing the system photos labeled as pineapples. Next time it comes across a spiky, knobbly ball in a fruit bowl, it should be able to identify it.

In unsupervised learning, a system is closely watched by its developers but can be taught to look for relationships in data by itself. You could show a system images of a range of fruits —without telling it which were which — and ask it to find similarities and differences between them. It might group together all the spiky, knobbly balls and suggest they are in fact the same thing, but it wouldn’t know they were pineapples unless it was told.

Unsurprisingly, the way we teach AI affects the way it learns.

Supervised learning methods are able to categorize and label data according to what humans already know, while unsupervised methods can be used for spotting patterns that people may not necessarily know to look for.