Start here
Your plan
| Days | Module | Slides |
|---|---|---|
| 1 | Setting up: data splits, bias and variance | C2 week 1 |
| 2 | Regularization: L2, weight decay, dropout and more | C2 week 1 |
| 3 | Normalizing, initialization, vanishing gradients, gradient checking | C2 week 1 |
| 4 | Mini-batches and exponentially weighted averages | C2 week 2 |
| 5 | Momentum, RMSprop, Adam, AdamW and learning-rate schedules | C2 week 2 |
| 6 | Hyperparameter tuning | C2 week 3 |
| 7 | Batch norm, softmax and frameworks | C2 week 3 |
| 8 | Experiments and interview drill | notebooks |
About 1 to 1.5 hours a day, for 8 days.
Code in this course is written for PyTorch 2.15 (October 2026), TensorFlow 2.21 and Keras 3.15. The slides use TensorFlow 1-style code in places; the notes show the current way.
How each lesson works: the slides for that part open at the top of the lesson. Read them, then the notes under them: the notes explain what the slide skips, show today's PyTorch and Keras code, and point to the experiments in the notebook. Notebooks are files that come with the course: open them in Google Colab (free, nothing to install), download them for Jupyter or VS Code, or read their PDF versions with all the outputs.
This is part 2 of 5: Improving deep neural networks. Part 1 built networks; this part makes them train well.
Credit: the slides are from the Deep Learning Specialization (course 2, Improving Deep Neural Networks) by DeepLearning.AI, taught by Andrew Ng, shared under Creative Commons BY-SA 2.0 for non-commercial educational use. Take the full course on Coursera for the videos and graded labs.
Maths is written in plain text: W·x is a matrix product, Σ a sum, ^ a power, [l] the layer, (i) the example.