Ask ten people what a neural network is and most will say "it's like a brain." True enough, and not much use when you sit down to learn how one works. A biological brain has billions of cells doing things researchers still argue about. A neural network in software is far more modest: a stack of numbers, multiplied and added in a particular order, arranged so that it gets good at spotting patterns. Once you can picture that shape, deep learning stops being mysterious and turns into detail.

A Neuron Is a Tiny Decision

You are standing at the door, deciding whether to take an umbrella. Three things run through your head: the sky looks grey, the forecast mentioned rain, and you hate arriving wet. Each factor pushes you toward yes or no, and some push harder than others. Then you decide.

A neuron in a network does the same job with numbers. It receives a few values, gives each one a certain amount of influence, adds everything up, and produces one number as its answer. That number might mean "definitely yes," "probably no," or something in between.

One extra step matters. Raw sums can be any size, which makes them awkward to combine. So the neuron passes its total through a small rule that squeezes the result into a predictable range, often between 0 and 1. Think of a referee who does not care how loudly you shouted, only whether you cleared the bar. In the field this rule is called an activation function, and it is the only jargon you need at the start.

Weights: Which Opinions Count Most

That "certain amount of influence" is a weight. A weight is a number that decides how loudly one input speaks. Picture a mixing desk with a row of volume knobs. Turn one up and that input shapes the outcome; turn it down and the neuron barely notices it.

Suppose a neuron is meant to help decide whether a photo shows a cat. One input might measure how pointy the ears look. Another might measure whisker-like lines. A third might report overall brightness, which is mostly irrelevant. Training will push the ear and whisker weights up and the brightness weight down, because that combination produces better answers.

Neurons also carry a bias — a baseline number added before anything else. Think of it as how easily the neuron says yes before it hears any evidence. A neuron with a strongly negative bias needs a lot of convincing.

Layers: Where Small Opinions Become Judgement

A single neuron can only draw a simple line through the data. Stack neurons in layers and something more interesting happens.

Input layer

The raw numbers go in. Pixels for an image, word counts for text, temperature readings for a sensor.

Hidden layers

These do the combining, and they work in stages. In image networks, the first hidden layer tends to respond to simple things like edges and patches of colour. A later layer combines those edges into shapes: a curve, a corner. A later layer still combines shapes into objects. Nobody programmed those stages. They emerged because that arrangement produced better answers.

Output layer

The final answer. A single number for "cat or not cat," or a set of numbers that add up to 1, one per possible category.

Depth is what puts the deep in deep learning. More layers mean the network can build more abstract ideas out of simpler ones.

How Learning Actually Happens

Nobody writes the weights by hand. They start as random numbers, which means an untrained network is confidently wrong about everything.

Training works in a loop. Show the network an example. Compare its answer with the correct one. Measure how far off it was. Then nudge every weight a tiny amount in whichever direction would have reduced that error. Repeat across thousands or millions of examples, and the weights drift into values that work.

Nothing inside the network knows what a cat is. The weights only record which combinations of pixels have tended to appear together.

The size of those nudges is the learning rate. Too small and training crawls; too large and it overshoots and never settles. If you have ever adjusted a shower tap, you already understand the trade-off.

What a Training Run Looks Like, Step by Step

1. Collect examples with known answers. Photographs already labelled, or house prices that actually sold.
2. Split them three ways. One set to train on, one to check progress during training, one kept untouched for the final test.
3. Push a batch through and collect predictions. Batches rather than single examples keep the arithmetic efficient.
4. Score the miss. A loss function turns "how wrong was that" into one number.
5. Send the error backwards. This is backpropagation: the error is used to work out how much each weight contributed, then each one is adjusted.
6. Repeat and watch the validation set. When validation performance stops improving while training performance keeps climbing, you have squeezed what this setup will give you.

Three traps that catch almost everyone

• Memorising instead of learning. Great scores on the training examples, disappointing results on anything new.
• A huge network on a tiny dataset. There is not enough evidence to fill all those weights sensibly.
• Optimising the wrong number. Impressive accuracy on a question nobody actually needed answered.

Your Next Move, Without Touching Code

Take a decision you make often, list the inputs you genuinely consider, and guess roughly how much each one matters. You have just described a neuron, a set of weights and a bias in one sitting.

Then, when you open your first tutorial, read the example slowly and ask three questions. What are the inputs? How many outputs are there? What is being measured as the error? Almost every model, from a two-layer toy to a network with hundreds of layers, can be unpacked with those three questions.

Resist the urge to memorise formulas first. The arithmetic is not the hard part; the shape is. Get comfortable with neurons as small decisions, weights as influence, layers as teamwork, and error as the thing that does the teaching. Everything you read afterwards will have somewhere to sit.

Photo: Kampus Production / Pexels