What learning from data means

A machine-learning model is an adjustable pattern.

Training uses examples to adjust that pattern. Evaluation asks whether the pattern still works on relevant examples it did not train on.

Why this matters

AI engineers rarely invent a learning algorithm from nothing. They define the problem, build evidence, select a model, fit it, evaluate it, and connect it to a reliable product.

If “learning” remains vague, it is easy to leak answers, optimise the wrong score, mistake memorisation for intelligence, or use reinforcement learning where a rule would be safer.

Plain-English mental model

Part 1 — Rules and learning solve different problems

A programmed rule states behaviour directly:

if attempts >= 5:
    require_identity_check()

The developer chose the threshold and action.

A learned model has adjustable values fitted from examples:

risk = model(attempts, device_age, location_change, ...)

The developer still chooses:

• the product goal;
• which data is legitimate;
• the model family;
• the objective and constraints;
• how success is evaluated;
• what happens when the model is uncertain;
• where human review and rules remain necessary.

Learning does not remove engineering decisions. It moves some behaviour from hand-written branches into fitted parameters.

When a rule is better

Prefer a clear rule when:

• the requirement is exact and stable;
• errors have severe consequences;
• there are too few representative examples;
• the decision must be directly auditable;
• the rule already performs the job cheaply;
• policy, law, or safety requires deterministic behaviour.

A good AI system often combines rules and models.

Part 2 — Examples, features, targets, and predictions

One learning example contains information about one case.

A featureFeatureAn input value presented to a model for one example, such as message length, account age, or a text embedding. A feature must be available at the real prediction time and calculated consistently. is an input value presented to the model.

A targetTargetThe answer a supervised model is trained to predict, such as a category, probability, or numerical value. Targets are also called labels in many classification tasks. is the desired answer in supervised learning.

For ticket urgency:

A label usually means a target category. Some teams use “label” for any target.

Features must exist at prediction time. The target must be defined consistently and represent the real goal closely enough to be useful.

If “urgent” means “an agent opened it within ten minutes”, the target may reflect staffing and queue position, not genuine urgency. Labels are measurements created by a process, not perfect truth.

Part 3 — Three kinds of learning feedback

The paradigms differ mainly in what feedback the learner receives.

Supervised learning

Supervised learningSupervised learningLearning a mapping from input features to known target answers using labelled examples. Classification predicts a category, while regression predicts a numerical value. uses examples paired with target answers.

Two common tasks:

• classification: predict a category, such as spam/not spam;
• regression: predict a number, such as resolution minutes.

It does not mean a human watches every training step. “Supervision” is the target signal in the training data.

Use supervised learning when historical examples contain a target that corresponds to the future prediction task.

Unsupervised learning

Unsupervised learningUnsupervised learningFinding structure in examples that do not include a desired target answer, such as groups, compressed representations, or unusual cases. The discovered structure still needs human interpretation and evaluation. receives examples without desired target answers.

It can search for:

• clusters of similar examples;
• lower-dimensional representations;
• density and unusual points;
• latent structure useful to another task.

A clustering algorithm may create three groups. It does not automatically know that the groups mean “billing”, “technical”, and “account”. A human or downstream evaluation supplies meaning.

“No labels” does not mean “no objective”. Algorithms still optimise something, such as compact clusters or reconstruction error.

Reinforcement learning

Reinforcement learningReinforcement learningLearning a policy for choosing actions through interaction with an environment and reward feedback, usually aiming to maximise expected cumulative reward rather than predict one fixed label., shortened to RL, learns action choices from interaction and reward.

Core terms:

• agent: the decision maker;
• environment: the world it interacts with;
• observation or state: information about the current situation;
• action: a choice the agent makes;
• reward: numerical feedback;
• policy: a rule or model mapping situations to actions;
• return: cumulative future reward.

Unlike ordinary supervised learning, a correct action label is not supplied for every situation. Actions change what happens next, and rewards can be delayed.

Exploration may affect real users. Train in simulation, constrain actions, use offline evidence, and require safety review where mistakes matter.

Not every AI problem belongs to only one box

A recommendation system might:

1. learn item representations without labels;
2. train a supervised click predictor;
3. optimise a sequence of recommendations with reward feedback;
4. apply deterministic safety and eligibility rules.

Name the feedback at each stage instead of labelling the whole product with one paradigm.

Part 4 — A model is a function with adjustable parameters

A model maps inputs to predictions.

For one feature:

prediction = weight × feature + bias

The weight and bias are model parametersModel parameterAn adjustable value learned from training data, such as a linear-model weight or a neural-network connection weight. Parameters change during fitting.. Training changes them.

A hyperparameterHyperparameterA setting chosen outside the parameter-fitting loop, such as learning rate, tree depth, or regularisation strength. It is normally selected with training and validation evidence, not the final test set. is chosen outside that parameter-fitting calculation.

Examples:

Hyperparameters are still informed by data, usually through validation. Do not choose them by repeatedly consulting the test set.

Model family and representation limit what can be learned

A straight line cannot represent every curve.

Features also shape the available pattern. If a model sees only message length, it cannot directly use message meaning.

Training finds parameters inside the chosen model family and representation. It does not search every imaginable program.

Part 5 — Loss turns prediction quality into a training signal

A loss functionLoss functionA numerical score describing how costly a model's prediction is relative to the training target. Fitting adjusts parameters to reduce an aggregate training loss, but lower training loss does not guarantee better real-world behaviour. assigns a numerical cost to a prediction relative to a target.

For numerical prediction, squared error is:

loss = (prediction - target)²

If the target is 10:

The square makes larger errors cost much more.

For classification, cross-entropy commonly scores predicted probabilities. Earlier you saw:

loss = -log(probability assigned to the correct class)

The training objective usually aggregates loss across a batch plus any regularisation.

Loss is a proxy, not the whole product

Lower loss does not automatically mean:

• fair decisions;
• useful user outcomes;
• safe behaviour;
• calibrated confidence;
• low latency or cost;
• correct behaviour under distribution shift.

Choose loss because it supports the product goal, then evaluate broader requirements separately.

Part 6 — The fitting loop

Fitting or training changes model parameters to reduce an objective on training examples.

For gradient-based fitting:

1. Read a batch of training features and targets.
2. Run the model to produce predictions. This is the forward pass.
3. Calculate loss.
4. Calculate how loss changes with each parameter: the gradient.
5. Use an optimiser to update parameters.
6. Repeat over batches.

One complete pass through the training dataset is an epoch.

A small fitting loop in Python

The next lesson derives linear regression carefully. Here, focus on the loop.

Expected output:

0 44.428
9 0.059
99 0.036
399 0.026
weight: 1.982
bias: 0.1
prediction for 6 hours: 11.99

This demonstration uses all five rows as training examples. A real evaluation must reserve unseen evidence.

Part 7 — Start with a baseline

A baseline modelBaseline modelA simple reference method that a more complex approach must beat to justify its added cost and risk, such as predicting the training mean or most frequent class. is a simple reference.

Examples:

• classification: always predict the most frequent training class;
• regression: always predict the training mean or median;
• forecasting: repeat the latest observed value;
• recommendation: show popular eligible items;
• language task: use a clear rule or existing retrieval method.

Why:

• proves the pipeline and metric can run end to end;
• reveals whether the task has useful signal;
• catches label or metric bugs;
• defines the minimum improvement needed;
• makes complexity earn its place.

from statistics import median

training_targets = [12, 14, 16, 100]
baseline = median(training_targets)

validation_targets = [13, 18, 40]
absolute_errors = [
    abs(baseline - target)
    for target in validation_targets
]

print("baseline:", baseline)
print("mean absolute error:", sum(absolute_errors) / len(absolute_errors))

Compare every candidate on the same validation cases and metric.

Part 8 — Training, validation, and test answer different questions

From the data-foundations lesson:

• training: what parameters fit the examples?
• validation: which model, features, hyperparameters, and threshold should we choose?
• test: after choices are frozen, how well does the selected system perform on untouched evidence?

The validation set indirectly influences the selected system. Repeated validation tuning can overfit validation too.

If experimentation is extensive:

• use cross-validation where appropriate;
• keep a final untouched test set;
• log experiments;
• limit manual metric hunting;
• obtain fresh future evidence after launch.

Do not merge the test set into training and continue reporting the old test score as independent evidence.

Part 9 — Generalisation is the real goal

GeneralisationGeneralisationA model's ability to perform usefully on relevant unseen cases drawn from the conditions it is intended to face, rather than merely remembering its training examples. means performing usefully on relevant unseen cases.

It is not merely “high test accuracy”. The test set must represent the intended use.

Memorisation versus useful pattern

Consider an image classifier:

• memorisation: recognises exact training image pixels;
• generalisation: recognises new relevant images under expected conditions.

Memorisation is not always useless. Retrieval systems deliberately remember records. The problem is claiming a model learned a reusable relationship when it only stored examples or shortcuts.

Why training performance can mislead

A highly flexible model can fit:

• real signal;
• random noise;
• duplicated examples;
• annotation quirks;
• watermarks, filenames, or source IDs;
• future information accidentally present.

Training loss usually improves as capacity and fitting time increase. Generalisation may improve, flatten, then worsen.

Distribution assumptions

Evaluation often assumes future cases are related to development data.

Independent and identically distributed, shortened to IID, is an idealisation: examples are independent draws from one fixed distribution.

Real systems violate it:

• users repeat;
• time changes behaviour;
• product policy changes labels;
• one model changes the data collected next;
• new languages, devices, or attacks appear.

Use group and time splits, stress tests, slice evaluation, and production monitoring to model those realities.

Distribution shift

Distribution shift means deployment data differs from training data.

Examples:

• new customer region;
• seasonal ticket topics;
• changed form fields;
• model-generated text entering the pipeline;
• a new fraud strategy;
• delayed or redefined targets.

A model can generalise within its test distribution and still fail after a shift. State the evaluation boundary honestly.

Part 10 — Overfitting and underfitting, first look

Underfitting means the model or training process cannot capture enough useful pattern.

Typical signal:

training error: high
validation error: high

Overfitting means the fitted system matches training-specific detail that does not transfer.

Typical signal:

training error: very low
validation error: much higher

These patterns are clues, not proof. A broken pipeline, leakage, label noise, or distribution mismatch can create similar curves.

Later lessons cover regularisation and the bias–variance trade-off in depth.

Worked example — route urgent support tickets

Goal: rank incoming tickets so genuinely urgent cases receive faster human attention.

1. Define the decision

At ticket creation, estimate whether the case meets a documented urgent policy.

The model assists queue ordering. It does not close cases automatically.

2. Define examples

One row is one ticket at creation time.

Features may include message text representation, declared issue category, and account service tier. Resolution outcome is unavailable at this moment.

The target is an urgency label produced from reviewed policy criteria, not merely whether an agent happened to answer quickly.

3. Select the paradigm

This is supervised classification: historical tickets have input features and target classes.

Unsupervised clustering might help explore topic structure, but clusters are not urgency labels.

Reinforcement learning is unnecessary for the first system. Online exploration that delays a real urgent ticket would be unsafe.

4. Split and baseline

Split by customer and time. Keep recent tickets as test evidence.

Baseline:

predict urgent only when a documented critical phrase matches

Also compare the most-frequent-class baseline, but inspect recall for urgent tickets; accuracy alone can hide failure on a rare class.

5. Fit and choose

Fit model parameters on training tickets. Use validation data to choose representation, regularisation, and operating threshold.

Do not train on validation labels after each decision.

6. Evaluate generalisation

On untouched test data, inspect:

• missed urgent cases;
• false alerts and workload;
• performance across language, source, and issue groups;
• confidence calibration;
• latency and fallback behaviour.

7. Monitor

Track input schema, category rates, delayed reviewed labels, overrides, and incidents. A shift triggers investigation and a separately evaluated model version.

In a real AI system

Common mistakes

Exercises

Knowledge check

Practice activity

Decision:
  rank a new ticket for human triage at creation time

Paradigm:
  supervised classification; reviewed tickets have policy-based labels

Example:
  one ticket snapshot at creation

Features:
  text representation, declared category, service tier

Target:
  reviewed urgent / not urgent under policy v3

Parameter:
  a fitted feature weight

Hyperparameter:
  regularisation strength

Split:
  customer groups kept together; train on older months,
  validate on the next month, test on the newest untouched month

Baseline:
  documented critical-phrase rule

Constraint:
  no automatic closure; uncertain high-impact cases go to a human

Generalisation boundary:
  expected to cover tickets from supported languages and current
  products under policy v3, not new products or unseen languages

Shortcut risk:
  agent queue ID may encode historical staffing rather than urgency

Key takeaways

Flashcards

One-page sketchnote summary

for batch in training_data:
  prediction = model(batch.features)
  loss = objective(prediction, batch.target)
  gradients = differentiate(loss)
  parameters -= learning_rate * gradients

Resources