So I built Art of DS — a data science and machine learning reference with 126 concepts, from statistics and metrics to transformers and diffusion models. Here is what you get from it.

Explanations you can trace back to the source

Every claim on the site points to the exact chapter or section where it is written — not just "see Goodfellow", but the section number, so you can open the book and read it yourself. There are more than 2,000 of these references across the site, drawn from around 170 books and papers.

The sources are the ones the field itself trusts:

• Dive into Deep Learning (Zhang, Lipton, Li, Smola)

• Deep Learning (Goodfellow, Bengio, Courville)

• Think Stats and Think Bayes (Downey)

• Introduction to Information Retrieval (Manning, Raghavan, Schütze)

• Forecasting: Principles and Practice (Hyndman & Athanasopoulos)

• the NIST/SEMATECH e-Handbook of Statistical Methods

• Google's Machine Learning Crash Course

• 84 research papers, from Attention Is All You Need and BERT to ResNet, GANs, CLIP and diffusion models

No random blogs, no "trust me". If you want to go deeper, the trail is already laid out for you.

Every claim points to a section of the book or paper that says it.

Visuals that explain, not decorate

Every page has an illustration, and about a hundred of them are animated: a thousand people showing why a test that catches 99% of the sick still flags mostly healthy people, a threshold sliding between two overlapping classes, gradient descent stepping down a loss curve, a perceptron passing a signal forward. Each figure shows one specific thing you can check against the text — the kind of picture that makes an idea click rather than fill space.

Bayes' theorem: 54 of 1,000 people test positive, but only 4 of them are sick.

Accuracy and its trade-offs: one threshold, two overlapping classes.

Every page is built the same way

Once you have read one page, you know where everything is on all of them:

1. Definition — one sentence and the formula, for when you just need to remember what something is.

2. Intuition — the idea in plain words.

3. Formal definition — the maths, done properly.

4. How it works — the mechanism, step by step.

5. Where it is used — real applications.

6. Why and when — when to reach for it, and when not to.

7. Worked example — a Python script that was actually run, with the output it really printed.

8. Pitfalls — the mistakes people make with it in practice.

9. Sources — everything the page relies on.

Short on time? Read the first two blocks. Preparing for an interview or a project? Read to the end.

Definition and formula first; the outline on the right is the same on every page.

A worked example: the output is what the script actually printed.

Tests that tell you what you actually know

Reading is not the same as understanding, so every page ends with three questions — 378 across the site. After each answer you see why it is right or wrong, which part of the page it checks, and which source it is based on.

Every answer is explained, with the section it tests and the source behind it.

When you feel ready, the final test (part of full access) draws twenty questions from across the whole site, one per page. Take it as a timed certificate attempt, score 70% or more, and you get a certificate — graded Good, Very good or Excellent — with a public page anyone can verify, a PDF and image download, and the details ready to add to LinkedIn.

Useful whether you are starting out or already working

If you are new, there are four learning tracks that take you through the concepts in the order they depend on each other — one of them goes from the train-test split all the way to transformers — plus a knowledge map that shows what you need to know before each concept. With full access, your progress, quiz results and the place where you stopped are saved across your devices.

The knowledge map: every concept, its topic, and how far it is from a starting point.

If you already work in the field, it is a fast reference. Press Ctrl K and search across about 650 terms and synonyms. Use the fourteen catalogues that compare whole families side by side — activation functions, optimisers, classification metrics, clustering methods, forecasting methods — and the pitfalls sections, which are often the most useful part for practitioners.

A catalogue: a whole family of methods, compared side by side.

A knowledge base that keeps growing

The glossary already holds a one-line definition for every concept, organised A to Z, and every page is available in both English and Ukrainian. The project is actively developed: I keep adding new concepts, terms and practice material, and recently updated pages are listed right on the home page. Everything I add later is included for people who already have access — you pay once and get the site as it grows.

Try it for free

Eighteen pages — at least two in every topic — are completely free, with their figures and quizzes, and need no sign-up. On every other page the definition, the formula, the intuition and the sources are open to everyone.

Full access is a single payment of $20, with no subscription, and a full refund within 14 days if it is not for you.

Good places to start: gradient descent, dropout, or the knowledge map to pick your own path.

If you spot something that could be explained better — or a reference that does not say what the page claims — I would love to hear about it at support@artofds.com.