Neural networks,
one idea at a time.

Combine inputs, weights, a bias, and an activation to make one prediction.

Add layers and watch simple boundaries combine into a shape that fits the data.

Compare ReLU, sigmoid, and tanh as the same signal moves through them.

Trace one value through every calculation in a tiny network.

Balance data, capacity, and noise while training and test scores diverge.

Turn prediction errors into a single signal the network can minimize.

Walk down an error landscape with controllable step size and momentum.

Send credit and blame backward through a network one connection at a time.

Move through a map where related words become nearby coordinates.

Inspect queries, keys, and values while a sentence resolves ambiguity.