LearnML Open LearnML

Learn how AI learns. Then read the papers that invented it.

Completely free. Short, playful lessons and labs you can poke, from your very first neuron all the way to real research papers, with an AI tutor that explains any line you don't get.

No maths or coding needed to start. Every doodle in this notebook works. Start with the neuron.

Start on page one See all 19 units

Should Nuri go to the park?

Park!

That's one artificial neuron: inputs times weights, added up, then a yes or no. Unit 6 builds one; Unit 7 wires thousands together and teaches them to set their own weights.

Draw a number. A real network will read it.

This is the digit reader from LearnML's Playroom, running right here in your browser. It's a small convolutional network: 5,994 weights that taught themselves from 60,000 handwritten digits.

On 10,000 digits it had never seen, it gets 98.5% right. See if your handwriting is in the other 1.5%.

In the app you can open it up: watch its filters find strokes, switch one off, and replay its training.

draw a digit here

It thinks it's

?

 

Its score for every digit, 0 to 9, rises here as you draw.

    One road, 19 stops, from zero to Transformers.

    Each unit is a few short lessons and a lab. Finish a unit and its landmark papers unlock. Tap a stop to see what's there.

      Then read the real thing, with help on every line.

      Each unit ends in the papers that started it: 56 landmarks, like the 1986 backpropagation paper and Adam, plus a library of about 900 more.

      • Select any sentence and ask: explain it simply, define the terms, give an example.
      • Tap a ✦ beside a heading, equation or figure to have that part explained.
      • Chat, Guide and Quiz for the whole paper, and your notes saved with it.
      • Circle anything on screen and the tutor explains what you circled.

      Try it on the paper here: tap a highlighted phrase.

      Nature, 1986 · Rumelhart, Hinton & Williams

      Learning representations by back-propagating errors

      We describe a learning procedure that keeps to shrink the gap between what the network outputs and what it should output. Along the way, the , which are neither inputs nor outputs, come to encode useful structure in the task, because the is carried backwards through each layer.

      Sample passage, paraphrased for this page.

      Pre-written example answers. In the app, the tutor answers your own questions live.

      Watch an AI learn to play. Then work out how it did it.

      The AI Game Lab turns famous game-playing agents into investigations: what the agent sees, what it's rewarded for, why it failed, and an experiment of your own.

      Pictures are real frames from each investigation, drawn by LearnML's own renderers.

      Why I built LearnML

      I wanted to learn machine learning: not just the buzzwords, but how it actually works.

      So I built a course that makes it easy. Every idea starts from zero, in plain words. You poke it in a lab, check you’ve got it, and move on one small step at a time, all the way to reading the real research papers, with a tutor to explain any line.

      It’s the course I wanted for myself, and it’s free for anyone who wants to learn the same way.

      The maker of LearnML

      Page one is waiting.

      Start with “What is machine learning?” It’s all free: every lesson, lab, paper and the AI tutor. Sign in with Google or GitHub, and your progress and notes follow you to every device.

      Start on page one