Degree 2 · Unit 2.2

How a machine learns

Inside every system that learns — from the spam filter on your email to the largest language model in the world — the same loop is running, millions of times over. It has four steps, and there is no fifth.

FIG. 7 — The training loop

It sees an example

An image, a sentence, a row of data — together with the right answer

It guesses

It produces an answer from its current state — entirely at random to begin with

The error is measured

How far its guess fell from the right answer — one number, called the "loss"

It adjusts itself slightly

In the direction that reduces the error — then back to the first step with a new example

↻

fawzooz.ai

Repeat that loop billions of times across trillions of examples, and what you get is what we call artificial intelligence today. There is no magic in it at all, only a vast amount of computational patience.

The three ways a model gains experience

Choosing how the model learns is the first strategic decision in any project — and the one that shapes the result most.

FIG. B5 — Supervised · unsupervised · reinforced

Supervised

Learns from labelled examples: "this plant is healthy", "this one is diseased".

Suits: classification and prediction · needs: labelled historical data

Unsupervised

Looks for similarity in unlabelled data and finds groupings nobody asked for.

Suits: customer segmentation and anomaly detection · needs: a great deal of raw data

Reinforced

Learns by trial and error: a penalty for a collision, a reward for getting closer.

Suits: sequential decisions and robotics · needs: an environment that can be simulated

A warning about data: most of a project's time goes to cleaning data, not to building the model. And biased data passes its bias to the model as a "rule of success".

fawzooz.ai

The right question is not "which algorithm is stronger?" but "which form of supervision does my data actually allow?". And most of a project's time goes to cleaning data, not to building the model.

Data governs everything

If the loop is the same in every system, what is it that separates a useful system from a harmful one? The answer is what the machine was shown. Data is not neutral fuel; it is the teacher. Every flaw in the data becomes a flaw in the system, except that it comes out polished, in confident language that hides where it came from.

Three flaws come up again and again. The first is absence: a group that is missing from the data will be the group the system always gets wrong. The second is bias: if the decisions of the past leaned one way, the system learns to repeat that lean and then presents it as objectivity. The third is the false shortcut, where the machine finds an easy marker that lands on the right answer without any understanding behind it.

Seven stages that turn and never end

AI without data is an engine without fuel — and this is the fuel cycle, whole.

FIG. B7 — DSLC, from raw data to continuous improvement

1 Acquiring the data

Internal and external sources and sensors, with a check on fitness.

2 Preparing the data

Cleaning, feature engineering, and a split into training, validation and test.

3 Modelling

Choosing the algorithm, tuning the hyperparameters, training.

↻

Iterations

Any stage may send you back to the one before

4 Evaluation

Performance measures, cross-validation, bias and fairness checks.

7 Improvement

New data, retraining, and realignment with the measures the business runs on.

6 Operation

Watching for data and model drift, with alerts and feedback.

5 Deployment

Batches, live inference or deployment at the edge, with versioning.

The path is a U: from 1 across to 3, then down to 4 and 5, then back across to 7 — and from there the cycle begins again.

fawzooz.ai

The cycle is not a straight line: evaluation often exposes a flaw in the data and sends you back to the second stage. That is the process succeeding, not failing.

Memorising versus learning

A student who memorises last year's exam questions does brilliantly on those questions and fails at everything else. A machine does exactly the same thing, and we call it overfitting: it masters the training data perfectly and then fails outside it. This is why no system is ever measured on the data it learned from, but on data it has never seen before.

You do not build models, so there is one practical sign to watch for: a striking performance in the demonstration, and a modest one in your actual working day. So when a vendor offers you a system with impressive accuracy, ask them a single question — what data was that figure measured on, and where did that data come from?

How a learning system perceives the world

A deep model does not see the image all at once; it builds it from the bottom up.

FIG. B9 — Hierarchical learning, layer by layer

Deep learning — layers that build understanding from the bottom up

No layer "understands" on its own: understanding is a property of the arrangement, not of the parts — which is why a model's decision cannot be explained by taking one cell apart.

fawzooz.ai

No single layer "understands" anything. Understanding is a property of the arrangement, not of the parts — which is why a model's decision cannot be explained by taking one cell apart.

FIG. 8 — Three states of a model

Underfitted

It has not learned enough — it errs in training and in reality alike

Balanced

It learned the general rule — it performs about as well inside its data as outside

Overfitted

It memorised the examples — excellent in training, weak in reality

fawzooz.ai

Do this

1 — On paper. Explain the four-step training loop to a member of your family, without one foreign term. If they do not understand, the fault is in your understanding, not in your explanation.
2 — On the tool. Ask an AI system to describe "the typical professional" in your field. Read the description critically: what bias did it inherit from its data? Which group is missing from its picture?
3 — In your field. Imagine a system learning from your organisation's decisions over the past five years. Write down three wrong patterns it would learn and repeat with confidence.

Where to after this unit? You now understand learning in general. The next unit goes into the kind you use every day: language models, and how predicting a single word manages to produce all of this.