What is the difference between supervised and unsupervised learning?

In supervised learning every training example carries the correct answer, so the model learns to reproduce a known mapping and can be measured against the truth. In unsupervised learning there are no answers, so the model looks for structure such as groups or anomalies, and evaluating the result requires human judgement because there is nothing to compare against.

Which machine learning algorithm should I use?

For a table of structured data, start with a gradient-boosted tree ensemble such as XGBoost or LightGBM, which is the strongest general default. Use logistic or linear regression when you need an explanation more than the last point of accuracy. Use deep learning when the input is raw text, images or audio. Choosing the algorithm matters far less than the quality of your features and your evaluation.

What is feature engineering and why does it matter so much?

Feature engineering is deciding what the model gets to see: which columns, in what form, combined how. It matters because a model can only exploit information present in its input, so a well-chosen feature that encodes domain knowledge often improves results more than any change of algorithm. On tabular problems it is usually where the largest gains come from.

How do I know if my model is any good?

Split your data before doing anything else, keep a test set you look at exactly once, choose a metric that matches the cost of each error rather than defaulting to accuracy, and compare against a trivial baseline such as always predicting the majority class. A model that does not beat the baseline by a meaningful margin has not demonstrated anything.

Why do models that pass every test fail in production?

The most common causes are data leakage inflating the offline score, a training set that does not resemble live traffic, features computed differently in production than during training, and the world drifting after training so the frozen model gradually describes a reality that no longer exists.

How much data do I need to start?

For a classical model on a table, a few hundred to a few thousand rows is a genuine starting point, provided the examples are representative and correctly labelled. Quality matters more than quantity: a thousand well-labelled rows beat fifty thousand where the labels were guessed.

Machine Learning: how it is organised, and how to know when it works

This course covers the discipline itself: how learning problems are categorised, which algorithm families exist and what each is for, why the features you choose matter more than the model you pick, and — most importantly — how to tell whether a result is real.

What this course sets out to do: give you the judgement to run a machine learning project. Choosing a problem type, choosing a metric, and recognising a result that is too good to be true.

What this course does not do: have you write the code. The premium catalogue does that, with notebooks and datasets.

The landscape

Course contents

#	Lesson	Main goal	Time
1	The three families of learning	Supervised, unsupervised, reinforcement — and which your problem is	9 min
2	The algorithms and when to use them	A practical map of what to reach for, and why	10 min
3	Features decide everything	Why the input matters more than the model	9 min
4	Evaluating honestly	Splits, cross-validation, baselines, and choosing a metric	10 min
5	Why models fail in production	The five failures that survive every offline test	9 min
6	Recap and FAQ	Synthesis, a project checklist, and 12 common questions	6 min
7	Quiz and attestation	Validate what you learned with 5 corrected questions	3 min

Is this course for you?

You have trained a model from a tutorial and want to know how to do it on your own problem.
You are an analyst or engineer being asked whether machine learning could help with something.
You manage a data team and want to ask better questions in reviews.
You keep hearing "the model gets 95%" and want to know what to check before believing it.

Recommended first: Introduction to AI, and ideally Python for AI.

What you will be able to do at the end

Classify a business problem as supervised, unsupervised, or not a machine learning problem at all.
Pick a sensible first algorithm with a reason you can state.
Explain why feature engineering usually beats algorithm shopping.
Set up an honest evaluation: splits, a baseline, a metric matched to the cost of errors.
Spot the signs of data leakage and of a score that is too good.
Anticipate the ways a model degrades after deployment.

Estimated time

Around 50 to 60 minutes of reading.

Prerequisites and next steps

Prerequisites: Introduction to AI for the vocabulary. Mathematics for AI makes the evaluation lesson land harder.

Natural continuation:

Deep Learning — when tables give way to images, text and audio
MLOps — keeping a deployed model alive
Ethics of AI — bias, explainability and what you owe the people affected

Frequent questions, answered in one line

Is machine learning always the right answer?

No, and the honest first question on any project is whether a rule would do. If your business logic is "flag any transaction over ten thousand from a new account", write that rule: it is deterministic, explainable, free to run and impossible to misinterpret. Machine learning earns its complexity when the pattern is genuinely too subtle or too varied to enumerate.

What is the difference between machine learning and statistics?

Substantial overlap, different emphasis. Statistics traditionally asks whether an effect is real and how confident we can be. Machine learning asks whether the prediction is accurate on data it has not seen. A statistician cares about the coefficient; a machine learning practitioner cares about the error on the test set. Both perspectives improve the other.

How long does a real machine learning project take?

Longer than the modelling suggests. In practice the split is roughly: understanding the problem and finding the data, a third; cleaning and building features, another third; modelling, a small slice; evaluating, deploying and monitoring, the rest. Teams that budget only for the modelling slice are the ones that miss deadlines.

Can I use machine learning with very little data?

Sometimes. On a table, a few hundred well-labelled rows can support a simple model. For text and images, fine-tuning a pre-trained model works with surprisingly few examples. What does not work is deep learning from scratch on a small dataset, which is a reliable way to produce a confident, memorising, useless model.

Want to build these models rather than read about them?

The premium catalogue walks you through real datasets, working notebooks and finished projects, with a verifiable certificate after a 40-question examination. Included in every paid plan.

Ready? Start with lesson 1 →

The landscape​

Course contents​

Is this course for you?​

What you will be able to do at the end​

Estimated time​

Prerequisites and next steps​

Frequent questions, answered in one line​