Learn deep learning by playing with it.
Every concept comes with a toy experiment you can touch. Search any idea, drag a slider, see why it works, and pass the challenge to turn it green.
Built with researchers from Tsinghua's College of AI. Toy experiments in the spirit of Physics of AI: simple enough to play, deep enough to build intuition.
Search any concept
A knowledge base of bite-sized Knowledge Units. Type “why does my loss explode” and land on the right one.
Play, don't read
Each unit has an interactive experiment that runs in your browser. No code, no setup, instant feedback.
Pass to turn green
Predict, experiment, then beat the challenge. Your map of deep learning fills up one green dot at a time.
Make your own course in one sentence.
Tell the AI Guide what you want to understand. It builds a path from reviewed Knowledge Units, in the right order.
What do you want to understand?
Gradient Descent
How a model walks downhill on its loss, one step at a time.
The learning rate is how big each step is. What happens if we make it very large, say 2.1?
Try 0.1, 1.0 and 2.1. Watch the dot and the loss counter.
The slope tells you which way is downhill. Gradient descent moves the weight a little in that direction, then checks the slope again.
The learning rate η sets the step size. Too small and you crawl. Too large and every step overshoots the bottom, landing higher on the other side, until the loss explodes. For this bowl, any η above 2 diverges.
Q1. Your training loss jumps around and then becomes NaN. What should you try first?
Q2. Lab challenge: get the loss below 0.01 in 5 steps or fewer.
Use the experiment above. Reset, pick a learning rate, and step.
Deep Learning 101
Course progress counts the units you passed. Finish all 16 to earn the DL 101 certificate.
Continue: Gradient DescentFrom a straight line to a Transformer. Follows the syllabus of Understanding Deep Learning, curated by Prof. Ziming Liu. About 4 hours of play.
Drag to reorder. Prerequisites missing from your course are flagged.
Tell me what you want to learn.
I find the right units or build a course from reviewed Knowledge Units. I never make up lessons.
New Knowledge Unit
Your unit appears here as the agent builds it.
My units
Your live units: 1,284 views · 71% pass rate
Course back office
Orange = below 50%, worth a second look from the creator.
What students want that we don't have yet.
| Unit | Creator | Submitted | Checks |
|---|
| Applicant | Background | Wants to write | Sample | |
|---|---|---|---|---|
| Chen Yu PhD student, Tsinghua | Optimization theory | Adam, Momentum, LR schedules | View draft | |
| Priya N. ML engineer | Computer vision | Convolution, ResNets | View draft |
| Course | Units | Learners | Completion | State | |
|---|---|---|---|---|---|
| Deep Learning 101 | 16 | 1,204 | 22% | Live | Edit order |
| Optimization in 1 Hour | 6 | 318 | 41% | Live | Edit order |
| Road to Transformers | 12 | — | — | Draft | Edit order |
| User | Role | Units passed | Credits used | Last active |
|---|---|---|---|---|
| Li Wei | Learner | 4 | 6 / 20 | Today |
| Wei Zhang | Creator | 15 | 12 / 20 | Today |
| Chen Yu | Applicant | 9 | 3 / 20 | Yesterday |
| Tsinghua AI-101 class | Group · 86 seats | 512 | 402 / 1,000 | Today |