The shape of the rest of this quest
The remaining seven tracks build deep learning the way you would actually learn it on a team: small concept first, then code, then operations.
- Neurons, Layers, and Forward Passes — what a neural network is, in shapes and matrix multiplies.
- Loss, Gradients, and Backpropagation — how a network actually learns, with calculus you can read.
- Training Data, State, and Checkpoints — the workflow scaffolding that turns "I trained a model" into "I trained, saved, and recovered a model".
- Stability, Regularization, and Making Training Work — the operational reality: most of training is making it not collapse.
- Architectures: Vision, Sequence, Attention — the named designs (CNNs, RNNs, Transformers) and why each was invented.
- Modern Deep Learning Practice — PyTorch mental model, transfer learning, mixed precision, experiment tracking.
- Choosing the Right Approach — the meta-skill of picking the right tool, including the choice not to use deep learning at all.
How to take this quest seriously
For each lesson, do the exercise. Run the code on your own hardware. Break it on purpose and read the error message — this is the single best way to internalize what each component does. The lessons are short on purpose; the depth comes from your hands typing the code.
Principle: This quest is a runway, not a destination. After you finish it, you should be able to read PyTorch source, follow a paper from arXiv to working code, and decide which model to fine-tune for your next problem.
Self-reference: The four canonical quests in this section — Pippa Quest, OO Quest, AI Math Quest, this one — together teach you the same skill from four angles: how a real system, a real abstraction, the math, and the model class fit together. None of them is a substitute for the others.