Of Layers Upon Layers, from Runway Unto the Clouds

Welcome back

We’ve now met every part of a single neuron. It takes in signals, multiplies each by a weight, adds them up with a bias, and passes the total through an activation function to decide what to send on.

But one neuron alone is like one termite: not much use. Today we’ll see how neurons are arranged together. To do that, let’s take a seat in the cockpit of an aeroplane.

How an aeroplane “thinks”

When a plane flies through cloud, information passes through it in stages.

First, dozens of sensors on the outside of the plane measure raw facts: air pressure, temperature, wind speed, tilt. Each sensor knows one tiny thing and nothing more.

Next, the instruments in the cockpit gather those raw readings and turn them into something more meaningful: altitude, airspeed, heading. An altimeter doesn’t just show “pressure”; it combines pressure readings into “we’re at 9,000 metres.”

Then the pilot looks across all those instruments together and forms a judgement: “We’re drifting slightly off course, and there’s turbulence ahead.”

Finally, a decision comes out: bank gently to the left.

Notice what happens at each stage. The information gets more meaningful as it travels. Raw numbers become readings, readings become a picture, and the picture becomes a choice.

Layers in a neural network

A neural network is built in exactly this way, in layers.

1. The input layer is like the sensors. It holds the raw data: the squares of a chessboard, the pixels of a photo, the words of a sentence. It doesn’t do any thinking; it simply passes the information inward.
2. The hidden layers are like the instruments and the pilot. Each neuron here listens to all the neurons in the layer before it, weighs their signals, and passes its own signal on. They’re called “hidden” simply because we don’t see them directly from the outside, just as passengers never see the instrument panel.
3. The output layer is the decision: “this is a cat,” “move the knight,” or “the next word is rain.”

A network might have one hidden layer or hundreds. When there are many, we call it deep, and that’s where the phrase deep learning comes from. It simply means a network with lots of layers.

The forward pass

When a network looks at something and gives an answer, the signal flows in one direction: from the input layer, through each hidden layer in turn, and out at the output. This journey is called the forward pass.

At every neuron along the way, the same small recipe from the last two lessons happens again: multiply by weights, add the bias, apply the activation function, pass it on. Nothing new is added. It’s the same humble step, repeated thousands of times, layer after layer.

And just like the aeroplane, each layer builds on the one before. In a network that recognises faces, the early layers tend to notice simple things like edges and patches of light. Middle layers combine those into shapes: an eye, the curve of a nose. Later layers combine those into “this is a face.” No one tells the network to organise itself this way. It tends to happen naturally as it learns.

What to carry with you