The foundation

The gate of this degreeOne page, in your words, explaining machine learning to a non-specialist.

You will never master a tool when you do not know how it works. This degree gives you the right mental model for it: how a machine learns, how language models "think", and why they are so often confidently wrong — and it does that without a single equation.

Entry condition

A coloured matrix and a signed contract from degree 1.

The gate

One page in your own words explaining machine learning to a non-specialist.

Units

Six

Dimensions

Knowledge · People

OPENING STORY

The model that succeeded for the wrong reason

On an early project, Fawzooz and his team built a system to tell machines infected with malware apart from clean ones. The result looked striking: accuracy above ninety-six per cent on the test data. He showed management the figure, the team was applauded, and people started talking about deployment.

Then they went back and examined why it had succeeded. It turned out that the infected machines in their data had all come from a single branch, and the logs from that branch carried a timestamp in a slightly different format from the rest. The system had learned nothing whatsoever about malware. What it had learned to read was the date format. Had they deployed it, it would have handed them ninety-six per cent of false reassurance, and the first machine from any other branch would have exposed it.

Fawzooz never forgot that lesson, and it is the first lesson of this degree: the machine does not learn what you want it to learn; it learns the shortest route to the right answer in your data.

Understand that sentence and you understand half of this subject. You will see why a model starts getting things wrong the moment it leaves the environment it was trained in, and why asking about a system's accuracy tells you nothing until you also ask which data that accuracy was measured on.

From data to predictions

The path a single data point travels until it becomes a decision — the production line this whole degree rests on.

FIG. B8 — The whole production line

Data

Raw information gathered as input.

⟶

Algorithm

A mathematical process that learns the patterns in the data.

⟶

New data

▼

A trained model

What comes of applying the algorithm to the data, used to predict.

⟶

Predictions

The model's output on new data: a decision or a classification.

New data enters at the trained model, not at the algorithm: training happens once and is repeated periodically, while inference runs every day.

New data enters at the trained model, not at the algorithm: training happens once and is repeated periodically, while inference runs every day. And the critical box is the last one: a model that does not end in a decision taken and measured is an expensive research project, not an asset.