What is Deep Learning? — one analogy

1The analogy

Crack open a water purifier and there isn't just one filter inside. Water comes in, and the first filter catches the coarse debris. The next one handles smell and color, the one after that catches finer particles still, and only after the last filter does drinkable water come out.

No single filter is impressive on its own. Ask the first one to make the water taste good and it can't — all it does is catch the coarse stuff. But when these simple jobs stack up one after another, the murky water that went in comes out clear at the end. No single filter, however fine its mesh, gets you there alone.

This is the structure behind deep learning. One layer does something very simple. Stack dozens or hundreds of those simple layers, and the early ones pick up on coarse features while the later ones catch increasingly fine ones. The number of layers is the depth — which is why it's called deep learning.

2In detail

What one layer does is simple

Open up a single layer and all you find is a simple calculation: multiply each incoming number by some fixed weight, add them up, and check whether the result crosses a line. It's simple enough to work out with pencil and paper.

The power shows up when layers stack. What the first layer produces becomes material for the second; what the second produces becomes material for the third. Just as water gets clearer with each filter it passes, a bundle of numbers changes shape a little with each layer it passes through.

So there's no need to wrestle with any single layer to understand deep learning. It's enough to picture something simple, repeated many times in sequence. This whole chain of connected layers is called a neural network.

Early layers stay coarse, later layers get fine

Look inside a deep learning system built to recognize photos and each layer responds to something different. The very first layer reacts only to something coarse, like the edge between light and dark. The next layer responds to the patterns those edges form together, and layers further along respond to the shapes those patterns build up into.

Only well toward the end does a signal for "what this actually is" appear — the same order as a filter catching the coarse stuff first and delivering drinkable water only at the very end. This shift from coarse to fine isn't something a person laid out in advance; it emerges on its own during training.

Nobody has to decide what to look at

Before deep learning, machine learning had one more job left for a person to do: deciding what to look at before making a judgment. To recognize a face, someone had to pick out things like the distance between the eyes or the length of the nose, turn those into numbers by hand, and the machine judged based on those numbers alone.

That step decided how well things worked, and it was also the biggest bottleneck — whatever clue a person missed, the machine could never see either. Deep learning hands that job over to the layers. Feed in the raw photo, and the early layers work out on their own what's worth looking at.

That cut one job for people and pushed performance up sharply, but it also made it much harder to know what the model actually looked at to reach its judgment. That's the reason deep learning models get called a black box nobody can see inside.

Going deeper raises the cost too

Adding layers widens the range of problems a system can handle, so for a long time people pushed to stack more of them. Simply piling them up didn't work well on its own, though — the deeper the stack, the fainter the learning signal got by the time it reached the earliest layers, leaving them barely learning anything. Only once several fixes for that problem arrived did truly deep networks become possible.

The amount of computation and the number of examples needed grow together too. More layers mean more values to tune, and more values to tune mean more examples are needed to tune them well. It's no coincidence that deep learning started working well right around the time graphics cards and large-scale data both became available.

So deep learning isn't always the right answer. When there are only a few hundred examples, or a decision needs to be explained, an older, shallower approach is often the better choice.

3More precisely

Deep learning is a branch of machine learning that trains an artificial neural network made of multiple layers on data. Each layer multiplies its incoming values by weights, adds them up, passes the result through a threshold called an activation function, and hands it to the next layer. Training measures how far off the final layer's answer is, then spreads that responsibility backward through the earlier layers, adjusting the weights as it goes. This backward pass is called backpropagation.

The water filter analogy breaks down in a few places. A filter cartridge comes ready-made from a factory, while a layer in deep learning starts with no criteria at all and works out what to filter for entirely during training. Water shrinks in volume as it passes through a filter, while a bundle of numbers passing through a layer can grow, shrink, or simply change shape. And a purifier lets you swap filters out one at a time, while deep learning's layers are trained together as one interlocking system — pull one out and swap in something else, and the whole thing usually stops working.

4Try it yourself

5Common misconceptions

It's easy to think deep learning copies the human brain directly, but actually it only borrows a loose idea from the brain, and the computation underneath is a simple repeated process that works very differently.
It's easy to think "deep" means difficult or profound, but actually it just refers to a structure with many stacked processing layers.
It's easy to think deep learning always beats machine learning, but actually when examples are scarce or a decision needs explaining, a shallower approach is often the better one.

7One-line summary

In shortDeep learning stacks simple layers so the early ones catch coarse features and the later ones catch fine detail, and deciding what to look at in the first place is now the model's job too.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02