From a neuron to a network

A single artificial neuron does something very simple: it multiplies each input by a weight, adds them up with a bias, and passes the result through an activation function.

output = activation(w1·x1 + w2·x2 + … + b)

On its own that is barely more than a linear model. The power comes from stacking many neurons into layers:

• The input layer takes the features — say, the eleven chemical measurements of a wine.
• One or more hidden layers each transform the output of the layer before.
• The output layer produces the prediction: one number for regression, or one score per class for classification.

Why activation functions matter

Without a non-linear activation, any stack of layers collapses into a single linear function. The most common choice today, ReLU, simply replaces negative values with zero: max(0, x). That small bend, repeated across thousands of neurons, lets a network approximate very complicated relationships.

"Deep" learning just means networks with many layers. Early layers tend to learn simple patterns (edges in an image, common word pairs), and later layers combine them into more abstract ones (faces, meanings).