What is deep learning?
In one line
Ages 8 to 10: Deep learning stacks many small steps, each building on the one before, to turn a picture or a sound into an answer.
Ages 11 to 13: Each layer works out weighted sums of what the layer before it found, and training nudges all the weights to make fewer mistakes.
Ages 14 to 18: A deep network learns its own features layer by layer, and backpropagation works out how every weight should change to reduce the error.
Deep learning is a kind of machine learning that uses neural networks with many layers. Each layer passes what it found to the next, so early layers pick up simple patterns and deeper layers combine them into bigger ones. The network learns by adjusting its numbers, called weights, from many examples.
Last reviewed: 20 October 2026 7 min read
The idea in one picture
Think of a relay of three people reading a handwritten number. The first person may only report small strokes: "a short curve here, another there". The second hears only those reports and combines them into shapes: "two loops, one on top of the other". The third hears only the shapes and decides: "that is an 8".
A deep network works in a similar relay, with numbers instead of words. Each layer receives the outputs of the layer before it and builds a slightly bigger idea. In a real network nobody tells the layers to look for strokes or loops. They find their own patterns during training, and those patterns are rarely as tidy as this example.
One small unit inside a layer, worked through: it receives 2 and 3 from the layer before, with weights 0.5 and 1. It computes 2 × 0.5 + 3 × 1 = 4, applies a simple rule, such as "if the total is below zero, use zero", and passes 4 on.
How it works, step by step
- Turn the input into numbers. A picture becomes pixel numbers; a sound becomes a long list of measurements.
- First layer. Many units each compute a weighted sum of the inputs and pass the result on.
- More layers. Each layer builds on the outputs of the one before.
- Answer. The last layer gives a score for each possible answer.
- Measure the error. The answer is compared with the label, the right answer for that example.
- Adjust. Backpropagation works out how each weight should change to reduce the error. This repeats over many examples and many rounds.
A little deeper (ages 11 to 13)
"Deep" counts layers. The number of layers between the input and the output is the depth of the network. A deep network is not deep in the sense of understanding more; it has more steps.
Layers multiply the number of weights quickly. If a layer of 100 units is connected to 100 inputs, every unit has one weight for each input: 100 × 100 = 10,000 weights, plus one bias per unit, 100 more. Finding good values for that many numbers takes many examples and a lot of computing.
Going further (ages 14 to 18)
Older machine learning often relied on people deciding which features to measure. Deep learning learns representations of the data directly, at several levels of abstraction. Between layers, each hidden unit applies a nonlinear activation function, such as the rectified linear unit (ReLU). Without it, stacking layers would add nothing, because a linear function of a linear function is still linear.
Different shapes of network suit different data: convolutional networks for images, video and audio, and recurrent networks and, today, transformers for sequences such as text and speech. The price of depth is that a trained network is hard to explain. Researchers can often interpret the first layer, but deeper weights are much harder to read. More layers also mean more weights to fit, so a deep network trained on too few examples can overfit.
Common mix-ups
- Mix-up: "Deep learning and AI are the same thing."Fact: AI is the whole field. Machine learning is one part of it, and deep learning is one kind of machine learning.
- Mix-up: "Deep means it understands deeply."Fact: Deep means many layers.
- Mix-up: "It is a copy of the brain."Fact: Each unit is a small calculation: multiply, add, apply a simple rule.
- Mix-up: "More layers are always better."Fact: More layers need more examples and computing, and can overfit.
Try it (no device, 5 minutes)
You need three people and a pencil. Person 1 secretly draws a simple object, such as a house, and describes only lines: "two slanted lines meeting at the top, four straight lines below". Person 2 hears only that and describes only shapes: "a triangle on top of a square". Person 3 hears only the shapes and guesses the object. Then swap jobs. Which layer lost the most information?
Quick check
What does the "deep" in deep learning mean?
Answer: The network has many layers, each building on the outputs of the one before.
The network has many layers, each building on the outputs of the one before.A unit receives 4 and 1 with weights 2 and −3. What is its weighted sum?
Answer: 4 × 2 + 1 × (−3) = 8 − 3 = 5.
4 × 2 + 1 × (−3) = 8 − 3 = 5.Why does each layer need a nonlinear activation function?
Answer: Without one, the layers together are still one linear function, so extra layers would add no power.
Without one, the layers together are still one linear function, so extra layers would add no power.Questions
Deep learning is a way for computers to learn from examples using neural networks with many layers. Each layer builds on what the layer before it found, so simple patterns are combined into bigger ones until the network gives an answer.
Machine learning is any method that lets a computer learn patterns from data. Deep learning is one kind of machine learning that uses neural networks with many layers and learns its own features from the data.
"Deep" refers to the number of layers in the neural network. More layers between the input and the output make a deeper network.
Deep learning is used in speech recognition, recognizing and finding objects in images, translating and generating text, and science such as genomics and drug discovery. It works best when there are many examples to learn from.
Only loosely. Each unit in a deep network is a small calculation that multiplies, adds and applies a simple rule, and the network learns by adjusting numbers, not the way a brain does.
Mechna Academy is a city of machines for curious children aged 8 to 18, opening first to ages 8 to 10. In each mission, a child guesses how a machine decides, then opens its Peek Hatch to check, changes one thing and tests it on a case it has never seen. Children learn how AI works this way, and they never chat with an AI.
For grown-ups: join the waitlistPlaces are limited. The waitlist is for parents and guardians 18 and older.