Machine learning, explained
How does a machine learn?
A scroll-driven walk, with no formulas: from a pile of messy data to a system that predicts — and knows when to doubt itself.
Scroll to begin
The data
It all starts with a pile of examples: photos, prices, messages, measurements. Each point in this cloud is a real case, with no order yet.
Cleaning up
Before learning anything, the house needs tidying: remove the duplicate or broken examples (in red), and set aside part of the data to train on and another part to check afterwards whether it really works.
The model tries
The system starts by guessing: it draws a simple boundary between the two groups. At first it gets many cases wrong — those are the blinking points, misclassified.
Learning
Every mistake nudges the internal parameters a little, in the direction that reduces it. The system repeats this thousands of times per second, and the boundary bends until it fits the data.
The prediction
A new, uncoloured case arrives — one the system has never seen. Learning is not memorising past examples: it is getting new ones right, with a probability attached to each answer.
Keeping watch
The world changes, and so does the data. A good system keeps watch over itself and raises an alarm when what it receives no longer looks like what it learned from.
This has been a simplified version, but the underlying process is this one: data, error, adjustment, prediction and monitoring.