Understand
the model.
Build the evidence.
Learn neural networks by building, breaking and diagnosing them.
The network receives numbers describing an image. No one gives it a rule for recognising a mug.
A simplified view of a trained network. Features are learned during training; this diagram illustrates the forward pass.
Six ways to make it stick.
Predict what will happen. Run the experiment. Explain what changed.
Read
Build the mental model.
Experiment
Change a variable. Watch the effect.
Run
Use real code and inspect the output.
Answer
Commit to a prediction or explanation.
Review
Retrieve it again before you forget.
Ask
Work through the part that did not click.
From intuition to engineering.
ML → DL
What changes when models learn representations from raw data?
3 lessons / 18 labs02PyTorch
How do I actually build a deep learning model?
4 lessons / 17 labs03Training dynamics
Why does a network learn well, badly, or not at all?
4 lessons / 14 labs04DL engineering
How do I run deep learning experiments professionally?
4 lessons / 8 labs05Representations
How do neural networks learn useful representations from different types of data?
6 lessons / 2 labs06Transformers
How does a Transformer process information and learn relationships between elements in a sequence?
6 lessons / 3 labs07Pretrained models
How can AI engineers leverage models that have already learned useful representations?
7 lessons / 3 labs08Fine-tuning
How can we adapt pretrained models efficiently and evaluate whether adaptation is worthwhile?
7 lessons / 4 labsBuild models you can explain.
- Trace a forward pass, a loss and backpropagation by hand, then build the same network in PyTorch.
- Diagnose training that stalls, overfits or explodes, and fix it with the right optimiser, schedule or regulariser.
- Run experiments you can trust and reproduce, with honest splits, seeds, checkpoints and tracking.
- Explain how embeddings and Transformers work, and adapt pretrained models with LoRA and QLoRA.
- Decide when a model is good enough to ship, and what it will cost to serve.
A few practical things.
Do I need an account?
No. Your progress, answers and notes stay in this browser. Turn on sync later if you want them on another device.
What do I need to install?
Nothing to start. Early lessons run Python in the page; every lesson also downloads as a notebook for Jupyter, VS Code or Google Colab.
How long does it take?
About 35 hours across ten modules, half or more of it hands-on. Lessons are written to be done one section at a time.
What if I get stuck?
Every answer is explained, the glossary defines every term, and an AI tutor beside each lesson can explain code or check your thinking.