About

A free course that teaches machine learning by making you run it. No lecture videos, no installation, no payment. Every algorithm is written from scratch and every number in the text is recomputed by a test before it ships.

123
lessons
401
steps
165
simulations
387
references

Who made this

Cagri Temel
Machine learning engineer · LLMs, retrieval and MLOps

I wrote every lesson, every simulation and every verification test in this course. The algorithms are not library calls: CART with Gini gain, random forests with bootstrap and feature subsampling, an MLP with backpropagation, a BPE tokenizer, skip-gram word2vec with negative sampling and the rest are implemented from scratch so that the numbers you see come out of code you can read.

I also wrote neural-trees, an open-source Python library for soft decision trees, along with the other models behind these lessons. One lesson in Track 1 trains a soft tree with it, so you can see the library do the thing the lesson just explained.

MS, Computer Science IEEE Senior Member AAAI Member Reviewer · IEEE SMC International Conference

Why this exists

Most courses explain an idea with slides, then hand you a notebook that already works. You watch someone else be competent. When the notebook breaks in your own project, none of what you watched helps, because you never made a decision.

This course inverts that. You get the controls first. You are asked to push a model until it fails, and then to say why it failed. The step does not open until you can.

That is also why there are no videos. A video runs at its own speed and cannot be interrogated. A simulation runs at yours.

How a lesson works

Every one of the 123 lessons follows the same five steps, so once you have done one you know what to expect from the rest:

01 Goal · what you are about to learn, in one sentence. 02 See · the algorithm runs in front of you with the active code line highlighted. 03 Play · the sliders are yours; break it. 04 Gate · either drive the simulation to a target or answer a question correctly. There are 154 of these gates and 129 questions. 05 Recap · the one thing worth remembering.

Wrong answers are not rejected. They are executed, so you can see what they actually do.

How the curriculum was designed

The five tracks were not chosen by topic popularity. Each one answers a question a learner actually gets stuck on, in the order they get stuck on it: what is an algorithm, how does a classical model make a decision, how does a network learn, how does a language model produce text, and how do you ship any of it without fooling yourself.

The last track is the one most courses skip. Knowing how attention works does not tell you how to build an eval set, debug a retrieval pipeline that returns confident nonsense, or estimate what a feature will cost before you build it. Those are lessons here.

How every number is verified

Courses quote figures. This one proves them. A verification script re-runs the algorithms and compares each result against what the lesson claims. If a single digit drifts, the build fails.

A number does not appear in a lesson unless a test recomputes it.

The current run checks 2,395 numeric claims and 0 of them disagree. It also proves all 154 unlock conditions are reachable by sweeping the parameter space, renders all 165 visualizations headless, and scans all 401 steps at real canvas dimensions for text that overflows or collides.

This has caught real mistakes: two key collisions that crashed lessons in the browser, a saturation metric computed on the wrong tensor, and five numbers left over from an earlier hyperparameter. The script is public, so you can run it yourself.

I wrote up the method and the worst thing it caught, a contrastive learning lesson that spent four paragraphs confidently explaining the behaviour of my own bug, in The Numbers in Your Tutorial Are Probably Wrong. Mine Were.

What you will know afterwards

These are drawn from the lessons that exist, not from a marketing wish list.

0
Start from Zero15 LESSONS
  • Trace an algorithm running step by step and say what it costs
  • Tell learning apart from memorising, and recognise overfitting from a curve
  • Explain why accuracy lies on imbalanced data, and pick a metric that does not
  • Split data without leaking, and catch leakage in someone else's setup
  • Decide whether one model is genuinely better than another using a statistical test
1
Classical Machine Learning37 LESSONS
  • Build k-NN, CART, random forests, boosting and SVM from scratch, and say when each fails
  • Read the bias–variance trade-off off a real curve instead of quoting it
  • Choose between L1 and L2 and explain what each does to the coefficients
  • Check whether a model's "70%" really means 70%, and fix it if it does not
  • Handle imbalanced data, distribution shift and time series without fooling yourself
2
Deep Learning22 LESSONS
  • Run backpropagation by hand on a small network and watch the gradients move
  • Compare SGD, momentum and Adam on the same problem and explain the difference
  • Say what batch norm, dropout, weight initialisation and clipping actually change
  • Read a convolution, a pooling layer and a residual connection
  • Describe how a diffusion model turns pure noise into a sample
3
Large Language Models29 LESSONS
  • Tokenise text yourself and explain why two tokenizers disagree on the same string
  • Compute attention by hand and assemble a full transformer block
  • Choose temperature, top-k and top-p deliberately instead of copying defaults
  • Explain why hallucination happens, and what retrieval does and does not fix
  • Adapt a model with LoRA instead of fine-tuning all of it
4
Using AI in Practice20 LESSONS
  • Build an eval set that catches a regression before your users do
  • Debug a retrieval pipeline that returns confident nonsense
  • Run a blind comparison between two models and read the result honestly
  • Estimate cost and latency before committing to a feature
  • Red-team your own system, then defend it

How this is paid for

It is not. There is no advertising, no sponsor, no investor, no paid tier and no payment integration of any kind. Nothing here is a funnel to something you buy later.

An account is free and exists for two reasons only: so your progress follows you between devices, and so you can leave a review. You can delete it, and everything in it, yourself in one step. What is collected and why is listed in the privacy policy, field by field.

Corrections

Pedagogy bugs matter as much as code bugs. If a step confused you, that is a defect worth reporting, and reporting it makes the lesson better for the next person. Open an issue on GitHub or write to contact@mltraining.org.