Neural networks, explained by analogy

Picture an enormous room of clerks arranged in rows. The first row looks at tiny fragments of a photo — just dots of light and dark. Each clerk shouts a small opinion to the next row: “I see an edge here.” The second row combines those shouts: “edges curving — might be an ear.” Row by row, opinions combine into bigger ones, until the final row announces: “cat.”

That is a neural network: layers of very simple units, each combining signals from the previous layer and passing its own signal forward.

The dials

Here’s the important part. Between every pair of clerks there’s a volume dial — how loudly should this clerk’s opinion reach that one? Those dials are called weights, and a large network has millions or billions of them.

Training is turning the dials. Show the network a photo of a cat. If it answers “dog”, an automatic procedure works backwards through the rows, nudging every dial a tiny amount in the direction that would have made “cat” slightly more likely. One example changes almost nothing. Millions of examples, millions of nudges — and the room of clerks becomes eerily good at cats, and dogs, and everything else it was shown.

What this picture buys you

• Nobody programs the middle rows. The useful internal opinions (“ear-like curve”) emerge from dial-turning. Even experts can’t always say what a given clerk learned — this is why you’ll hear AI called a “black box”.
• Ability is stored as numbers. A trained model is, literally, a giant list of dial settings. Copy the numbers, copy the ability.
• It’s still pattern-matching. Bigger rooms with more clerks capture subtler patterns — but lesson 3’s warning stands: patterns, not understanding.

Check your understanding

1. During training, what changes inside a neural network?

The number of layers grows with each example The weights are nudged slightly after each example Nothing — networks are fixed at creation

2. Why are neural networks sometimes called 'black boxes'?

They are stored in sealed hardware Their useful internal patterns emerge from training and can be hard for anyone to interpret Their code is always secret

Check answers

Recap

Layers of simple units, connected by millions of adjustable dials, tuned by examples. That’s the engine. Next in E1: what makes an AI “generative”, and what a large language model really does when it talks to you.