Neural networks, from first principles
A hand-held tour through the linear algebra, calculus, and engineering that make a neural network actually learn. Every chapter has runnable demos right next to the math — no faith required.
How to use this guide
Read the chapters in order if you're new — each one builds on the previous. Every concept appears in three forms: the equation (so you can be precise), an explanation in plain language (so you can build intuition), and an interactive demo (so you can poke it until it clicks).
When a paragraph references something earlier, it links back to it. When a chapter introduces a new symbol, it's defined the moment it appears. Bring a pen — writing the math out helps.
The map
Vectors, dot products, matrix multiplication — the language of every neural network.
Derivatives, partials, gradients, the chain rule. No epsilons, no measure theory.
From a single neuron to a deep stack. Forward pass derived end to end.
How a network actually learns. Step sizes, momentum, and what Adam is doing.
The algorithm — derived from scratch, line by line, matching the code in this repo.
Open the workspace and watch your network learn the very things you just studied.
Skip ahead to chapter 5 or jump straight into the interactive workspace. You can come back to fundamentals when something stops making sense.