Concepts

Deep learning

4 min readbeginnerUpdated 28 Sept 2026
1 · In one line

Deep learning is machine learning with neural networks that stack many layers, so each layer can build a more abstract picture of the data than the one before.

1 · What it is

Deep learning is a branch of machine learning that uses neural networks with more than one hidden layer, and usually many. Data goes in at one end, passes through the layers in order, and a prediction comes out at the other end.

The stacking is the point. Each layer takes the description produced by the layer before and turns it into something more abstract. In a network trained on photos, the first layer might pick out edges and corners, middle layers join those into shapes or parts, and the last layers react to whole objects.

No one writes those descriptions by hand. During training, the network compares its prediction with the right answer and passes the error back through every layer, nudging each internal number, called a weight, so the next prediction is a little better. Once trained, the same weights are used to make predictions on new data.

The idea is decades old, but it took fast graphics chips and very large datasets to make it work at scale. In 2012, Hinton and his students built a network with 60 million parameters that sorted photos into 1,000 categories. In the ImageNet object-recognition contest it cut the error rate nearly in half, and computer vision research changed course around it.

2 · Why it exists

Some patterns are too tangled to describe by hand.

Straight lines fall shortA simple model that weighs each input separately cannot separate data whose pattern curves or wraps around itself.
Hand-built featuresBefore deep networks, people often had to invent combinations of inputs by trial and error to expose the pattern.
Raw pixelsA photo reaches the computer as a grid of pixel values. Something still has to connect those numbers to an answer such as “cat”.
3 · How it works

Follow one input through a deep network.

STACK · HIDDEN LAYERS 1 TO 3 each layer builds on the output of the one before INPUT Pixels raw numbers LAYER 1 Simple patterns edges, corners LAYER 2 Shapes combined edges LAYER 3 Objects whole things OUTPUT Prediction “dog” PREDICTING: DATA FLOWS FORWARD, LAYER BY LAYER TRAINING: THE ERROR FLOWS BACK AND EVERY WEIGHT IS ADJUSTED
  1. 1 · inputThe raw values, such as the pixels of a photo, enter the first layer of the network.
  2. 2 · weighEach neuron multiplies its inputs by learned weights and adds the results together.
  3. 3 · bendAn activation function reshapes that sum so the network can capture curved, non-linear patterns.
  4. 4 · stackEach layer passes its output to the next, so every layer works with a more abstract version of the data.
  5. 5 · correctDuring training, the error on each prediction is sent backwards through the layers to adjust every weight.

Deep simply means more than one hidden layer. The depth is what lets each layer build on the last.

4 · Where it's used
WhoWhat they askWhat it works with
Photo library team“Which of these pictures show a dog?”Pixel values of each image
Voice assistant team“What did the person just say?”The recorded sound, as a stream of numbers
Translation service“How does this sentence read in Spanish?”The sentence as a sequence of words
Genomics lab“Which changes in this DNA sequence might matter?”Long genetic sequences
5 · What it solves, and what it doesn't
solves
  • It learns useful combinations of inputs by itself, instead of relying on people to design them.
  • It can model relationships far more complex than a straight line.
  • It brought large accuracy gains in recognising speech and objects in images.
doesn't solve
  • It depends on large datasets to find its patterns.
  • Its reasoning is hard to read; unlike a simple linear model, you cannot just inspect its weights.
  • Doing well on training data does not guarantee correct answers on new, unfamiliar data.
  • It is computationally heavy, which is why specialised chips exist to train and run it.
6 · Go deeper

Sources used

This explainer is written in original language. The links below support its factual claims.

  1. paperDeep learning, Nature (LeCun, Bengio and Hinton, 2015) · read 27 Sept 2026
  2. docsMachine Learning Glossary, Google for Developers · read 27 Sept 2026
  3. docsNeural networks (Machine Learning Crash Course), Google for Developers · read 27 Sept 2026
  4. docsNeural networks: Activation functions (Machine Learning Crash Course), Google for Developers · read 27 Sept 2026
  5. paperImageNet Classification with Deep Convolutional Neural Networks, NeurIPS 2012 (Krizhevsky, Sutskever and Hinton) · read 27 Sept 2026
  6. officialFathers of the Deep Learning Revolution Receive ACM A.M. Turing Award, ACM · read 27 Sept 2026
  7. officialInceptionism: Going Deeper into Neural Networks, Google Research · read 27 Sept 2026