01 / 06 · Scene
Linear regressionData science, explained from the source.
Every concept explained from the book or paper it comes from, with a figure and a worked example that was actually run.
Read a free page Search concepts, terms, definitions…Ctrl K
18 pages free in full. The rest unlocks once for $20, with no subscription.
Free account · 18 free · 108 lockedUnlock
Lifetime access · every page is open to youBrowse all concepts
In numbers
- 126concepts, from expectation to the transformer
- 378quiz questions, each answer explained
- 266books, papers and docs cited to the section
- 18pages free to read in full, with no account
Your place, kept
What you have read, how your quizzes went, and where to pick up — recorded as you read.
Loading your progress
What is free
What a locked page still shows
A locked page is never a dead end. The definition, the formula and the intuition are open to everyone; the rest of the page opens with access.
Open to everyone
- Definition
- Intuition
- Sources
With access
- Formal definition
- How it works
- Where it’s used
- Why & when
- Worked example
- Pitfalls
The locked text is never sent to your browser: not in the page, not in search, not in the feed.
Sigmoid Locked01 · Definition Turn a score into a smooth value between zero and one, useful for binary probabilities and gates.02 · Intuition 6 sections · With access09 · Sources Dive into Deep LearningDeep Learning 2016
Contents
Contents
Foundations first, applications last. Every topic has free pages you can read in full.
Mathematics4 pages 2 free Statistics15 pages 2 free Metrics & losses21 pages 2 free Optimization8 pages 2 free Machine learning30 pages 2 free Deep learning16 pages 2 free Computer vision13 pages 2 free NLP17 pages 2 free Tooling2 pages 2 free
Proof
Why trust it
- 01
Dive into Deep LearningZhang, Lipton, Li, Smola
§12.4.1§12.5§12.3.1.1
Deep LearningGoodfellow, Bengio, Courville · 2016
§8.2.2–8.2.3§4.3.1
Cited to the section
Every claim points to the chapter or section of the book or paper that says it. The full list is public.
Audit the sources - 02
Gradient descent on a narrow valley — the path crosses the floor far faster than it runs along it. A figure that argues
Each page has a figure that makes a point you can check — not a decoration.
See an example page - 03
$ python gradient_descent.pyeta=0.1 w= 2.7938 loss=0.0425eta=0.5 w= 3.0000 loss=0eta=0.9 w= 2.7938 loss=0.0425eta=1.05 w= -6.4153 loss=88.65Examples that were run
Worked examples are real scripts, and the output shown is what they printed. When a result disagreed with the textbook, the page says so.
How it is made
Sources
- Dive into Deep Learning
- Deep Learning2016
- Think Stats
- Introduction to Information Retrieval2008
- Deep Residual Learning for Image Recognition2015
- The Curious Case of Neural Text Degeneration2020
- An Image is Worth 16x16 Words2021
- BERT2019
- Denoising Diffusion Probabilistic Models2020
- Language Models are Few-Shot Learners2020
- Learning Transferable Visual Models From Natural Language Supervision2021
- On Calibration of Modern Neural Networks2017
Tracks
Tracks
A track is an ordered path through the pages, each with its quiz. With lifetime access, its ring fills as you read them.
Data science from zero
From a train/test split to transformers, in the order each page needs.
0 of 12 read 12 pages 5 freeStart the trackStatistics you actually need
Distributions, inference, and the mistakes that sink analyses.
0 of 15 read 15 pages 2 freeStart the trackMachine-learning models explained
Linear models, trees, ensembles, clustering — how each one fails.
0 of 30 read 30 pages 2 freeStart the trackDeep learning, end to end
Networks, optimisation, regularisation and what to do when training fails.
0 of 16 read 16 pages 2 freeStart the track
Tracks are being built; these four start from pages that already exist.
Practice
Check yourself
Every page ends with three questions. Each answer is explained, with the section it tests and the source behind it.
Free sample Gradient descent
3 questions
Each answer is explained, with the section it tests and its source.
After each answer
- Why it is right or wrong
- The section it tests
- The source behind it
The final test
Twenty questions drawn at random from every page on the site, one question per page so no corner of the site dominates. Nothing is shown until you finish; then every answer is explained. It comes with lifetime access.
Take the final test Unlock the final testAccess
Read the start free. Unlock the rest once.
A few pages in every topic are free. One payment unlocks everything else — every page, every quiz, your progress on all your devices — and nothing renews.
Lifetime
$20 one time
Pay once, keep access
- Everything in Free, every cheat sheet and its PDF included
- All pages, in full
- The quiz on every page, and a final test drawn from all of them, saved to your account
- Progress and tracks in your account, on all your devices
- New pages and practice as they land
You have access
Paid once. No renewal and no charge to come.
Free $0 No account needed
- Free pages, in full
- Definition, formula and intuition of every page
- Quizzes on the free pages
- Every cheat sheet; the PDF with a free account
- Every catalogue
Prices are in US dollars. Any VAT or tax that applies where you live is added at checkout. You can read the Terms and the Refund policy before you pay. Terms · Refunds
| Free | Lifetime | |
|---|---|---|
| Free pages, in full | Included | Included |
| Definition, formula and intuition of every page | Included | Included |
| Quizzes on the free pages | Included | Included |
| Every cheat sheet; the PDF with a free account | Included | Included |
| Every catalogue | Included | Included |
| All pages, in full | Not included | Included |
| The quiz on every page, and a final test drawn from all of them, saved to your account | Not included | Included |
| Progress and tracks in your account, on all your devices | Not included | Included |
| New pages and practice as they land | Not included | Included |
Catalogues
Catalogues
Families of methods, compared side by side.
- Deep learningFreeActivation functionsNonlinear activations give composed affine layers more expressive power. Compare their ranges and local derivatives without mistaking those derivatives for a whole network’s gradients.5 entries
- Metrics & lossesFreeClassification metricsFive scores summarize confusion-matrix counts at a decision rule; ROC AUC summarizes score rankings. Log loss and calibration assess probabilities. Their blind spots differ.8 entries
- Machine learningFreeClustering methodsClustering objectives and geometries side by side: centres, densities, mixtures and hierarchies. These are conditional modelling choices, not a ladder of universal replacements.7 entries
- Computer visionFreeComputer vision architecturesThe lineage of image-recognition networks from AlexNet to the Vision Transformer, and what each one changed about how images get processed.5 entries
- StatisticsFreeForecasting methodsEvery row is a different answer to what the future should look like. The first one has no parameters at all, and more methods fail to beat it than anybody reports.8 entries
- Machine learningFreeLinear model familyLinear-model objectives and penalties compared. Coefficient linearity permits nonlinear input features; uncertainty and tuning depend on the objective, normalization and modelling assumptions.7 entries
Recently updated
Pages added or revised lately, newest first.
- BackpropagationWork backwards from a network’s error to find how every weight contributed, reusing the work instead of testing weights one by one.
- BaggingFit models on bootstrap versions of the data and combine their predictions to make an unstable learner less sensitive to one sample.
- Batch normalizationRecenter and rescale a layer’s activations during training so their changing scale is easier to work with.
- Bias-variance tradeoffSeparate a model’s systematic misses from its sensitivity to the training sample to understand prediction error.
- CLIPLearn a shared image–text space so descriptions can retrieve images or define a classifier’s candidate labels.
- ConvolutionSlide the same small filter across an image to find local patterns wherever they appear.
Questions
Questions
What is free?
A few pages in every topic, read in full with no account. On every other page the definition, formula, intuition and sources are open too.
What does the one-time purchase include?
Every page, catalogue and quiz on the site, now and as it grows, with no renewal. Your progress and quiz results are kept in your account and follow you to every device.
Do I need an account?
Not to read the free pages. You need one to buy, and lifetime access keeps your progress and quiz results in your account, on every device. You sign in with a link or a code from an email; there is no password.
Is the locked text hidden, or just blurred?
It is not sent to your browser at all, so there is nothing to find in the page source, the search index or the feed.
Can I get my money back?
The Terms and the Refund policy are shown before you pay. It is one payment, not a subscription, so nothing renews and there is nothing to cancel.
Does the site track me?
No ads and no tracking cookies. Visits are counted anonymously with Umami: no cookie, nothing that could tell who you are, and nothing at all from a browser that asks sites not to track it. Signing in sets the cookies that keep you signed in: the session itself; one that remembers whether your account is free or has access, so pages draw right; and, for the hour it takes to finish signing in, one holding the address the code went to. The cookies page lists them.
Read a page before you decide.
The free pages are free in full, with no account. If they are what you were looking for, one payment unlocks the rest.
Paid once. No renewal and no charge to come.