Neural Networks & Deep Learning
How AI actually learns — the maths-free explanation with real analogies
Contents
The brain analogy (and why it is imperfect)
Neural networks are loosely inspired by the human brain — but calling them "artificial brains" is a bit like calling a bicycle an "artificial horse." The idea is borrowed, but the mechanism is completely different.
A neural network is a mathematical function made of layers of connected numbers. Data goes in one end, gets transformed through many layers of arithmetic, and a prediction comes out the other end. That is genuinely all it is.
The "neural" part comes from how the connections are structured — in loose analogy to neurons and synapses. But unlike real neurons, artificial ones just multiply numbers and add them up.
Layers: how information flows
Every neural network has three types of layers:
Input layer: Raw data enters here. For an image, this is every pixel's colour value. For text, this is the numerical representation of each word.
Hidden layers: The "thinking" happens here. Each layer learns to recognise increasingly abstract patterns. Early layers might detect edges in an image. Middle layers detect shapes. Later layers detect concepts like "face" or "car".
Output layer: The final answer. For image classification, this might be 1000 numbers — one per category — each representing the probability of "this is a cat" or "this is a dog".
Deep learning simply means there are many hidden layers — sometimes hundreds. More layers = ability to learn more complex patterns.
Training: how a network learns
Training is the process of adjusting all those numbers (called weights) until the network makes accurate predictions.
The loop works like this: 1. Show the network an example (say, a photo of a cat) 2. The network makes a guess ("I think this is a dog — 73% confident") 3. Calculate how wrong it was (the "loss") 4. Backpropagation: work backwards through the network, nudging every weight slightly in the direction that reduces the error 5. Repeat millions of times
After enough examples, the network has tuned its weights to correctly classify cats, dogs, and everything else it was trained on. This tuning process is called gradient descent.
Why deep learning changed everything
Before deep learning, AI required humans to manually define the features to look for — "check if there are whiskers, pointed ears, and fur." This was slow, expensive, and fragile.
Deep learning learns the features automatically from examples. Given enough labelled photos, it figures out what a cat looks like without anyone defining "whiskers" or "fur."
This is why deep learning exploded when large datasets and powerful GPUs became available. The algorithm itself was known since the 1980s — what changed was scale.
Transformers (the architecture behind GPT, Claude, and Gemini) are a type of deep neural network optimised for sequence data like text. They use a mechanism called "attention" to weigh which parts of the input matter most for each prediction.