The course labs are an Education feature, coming in a later release: about 12 labs for a first deep-learning course, each with a teacher’s pack. This tutorial plans a term with the ten lessons every edition has today, and says what the course labs will add. The lessons stay free for students, in every edition.
The ten lessons, in order
Each lesson is a lab: a starter, a worksheet, locks and checks. Each trains on a laptop’s CPU in under a minute.
- Two moons: a hidden layer and an activation, so a classifier can bend its boundary.
- MNIST digits: from a Linear classifier of pixels to a small convolutional network.
- Regularization: a large network memorizes 2,000 pictures; Dropout fixes it.
- Residual networks: skip connections, so a deep stack of convolutions learns.
- Attention: a transformer encoder layer tells which digit comes first.
- A transformer from scratch: an encoder and a decoder from one-line blocks.
- Predicting a number: a regression, with its own loss and error metric.
- Remembering order: an LSTM, then a GRU and an RNN, on digits in order.
- Training that behaves: gradient clipping, then early stopping before noisy labels.
- A language model: a small GPT on Tiny Shakespeare, measured by its perplexity.
A term, week by week
A 12-week term with one lab a week can run:
- Weeks 1 and 2: Lessons 1 and 2, the blocks, the Lab page and Check my work.
- Week 3: Lesson 3, overfitting and regularization.
- Week 4: Lesson 7, regression, so students see a second kind of loss early.
- Weeks 5 and 6: Lesson 4, then a lab of your own on convolutions (Writing a lab).
- Week 7: Lesson 9, training that behaves.
- Week 8: Lesson 8, sequences.
- Weeks 9 and 10: Lessons 5 and 6, attention and transformers.
- Week 11: Lesson 10, a language model.
- Week 12: a project of the student’s own choosing, from an example or their own data.
Lessons 4 and 10 download their datasets. In a lab whose network is blocked, have IT put them in place before those weeks (Running a lab).
Each week
- Before: do the lab yourself, and read its worksheet’s questions.
- In the lab: students work through the steps and press Check my work as they go.
- After: grade the hand-ins (Grading a class), and start the next class with the summary’s most common error.
What the course labs will add
The course labs follow the lessons where they fit, and add labs of their own for what a course needs beyond them: data splits and overfitting, convolution in depth, transfer learning, and a project. Each comes with a teacher’s pack:
- the student’s file, ready for your LMS;
- the worksheet and the solution;
- hidden checks, which mark the work without being in the student’s file;
- a marking guide, the time the lab takes, and notes on the mistakes students make most.
The packs reach your school with its Education license, never inside the app, so no student can read a solution there.