Neural Network Tutor: From Replay Lab to Production Debugger

Suppose we want to teach a tiny robot whether a fruit is an apple. Before drawing a neuron, we must answer simpler questions. What is one fruit? Which clues may the robot inspect? Where is the answer stored? How do we prevent it from reading that answer before making its guess?

Only then do we build the first neuron. Roundness and redness cast weighted votes. We add those votes, include a starting preference called bias, and make one prediction. A neural-network name is not permission to leave an idea mysterious.

Then we deliberately break that tiny model.

It fails on XOR. It loses information when a layer becomes too narrow. It memorizes when a network becomes too large for its evidence. Its gradients vanish, its ReLUs die, its attention maps look persuasive without proving causality, and its validation performance can change long after its training accuracy becomes perfect.

Every later idea must repair a failure that we can already see:

perceptron
    ↓ one boundary is not enough
hidden representation
    ↓ shape controls information flow
width / depth / expansion / compression / skips
    ↓ data has structure
CNN / RNN / LSTM / encoder-decoder / attention / transformer
    ↓ a prediction is still not an explanation
activation / gradient / representation / ablation inspection
    ↓ laboratory behavior must survive real execution
replay trace / run comparison / bounded production debugger

This is not a gallery of network pictures. The course is organized into 47 stories covering 102 named concepts. Closely connected ideas share a story when one naturally creates the need for the next. Every concept remains directly linkable and keeps its own explanation, visual scene, use case, limitation, and safe conclusion.

The order inside every story is deliberate:

what we already understand
        ↓
the exact thing that breaks
        ↓
a child-sized analogy—and where it stops being true
        ↓
a tiny example with real numbers
        ↓
predict what the animation should do
        ↓
run the narrated visual
        ↓
calculate it by hand
        ↓
use it, diagnose it, and compare it

Small mechanisms are computed in the browser. Long-running or research-scale behavior is shown as a deterministic trace or an explicitly labelled conceptual reconstruction.