# What is Deep Learning?

Deep learning is a subfield of machine learning that uses neural networks with many layers of processing units to learn representations directly from raw data. Each layer transforms the output of the previous one, allowing later layers to encode increasingly abstract features. This layered structure underpins most modern speech, image, and language systems.

The word deep refers to depth of composition rather than sophistication. Early layers of a vision network may respond to edges and textures, middle layers to parts, and later layers to whole objects. Because these features are learned rather than hand designed, deep networks removed much of the manual feature engineering that dominated earlier practice in vision and speech.

Training relies on backpropagation, which computes how much each parameter contributed to the error and adjusts it accordingly, combined with gradient descent over batches of examples. Progress has depended heavily on three external factors, namely large datasets, parallel hardware such as graphics processors, and architectural advances including convolutional networks, recurrent networks, and the transformer.

Depth brings costs. Deep networks require large amounts of data and computation, are sensitive to initialization and hyperparameters, and offer limited insight into why a particular output was produced. Interpretability research addresses the last point, but explanations remain partial. Deep models can also absorb and amplify statistical bias present in the corpora they were trained on.

Nearly all contemporary large language models are deep networks, generally transformer based, trained on very large text collections. The same family of methods also covers image classification, speech recognition, protein structure prediction, and recommendation. What unites these applications is the use of stacked differentiable layers optimized end to end against a single training objective.

## Key points

- Neural networks with many stacked layers of learned transformations.
- Learns features automatically instead of using hand engineered ones.
- Trained by backpropagation and gradient descent.
- Needs large datasets and substantial computation.
- Underlies modern language, vision, and speech systems.

## In practice

An image classifier trained on photographs illustrates the layering. The first convolutional layers respond to simple gradients and edges. Deeper layers combine those signals into shapes such as wheels or eyes, and the final layers map combinations of parts onto category labels. No one specified an edge detector, since every filter was learned from labeled images during training.

## Related terms

- [Neural Network](/en/glossary/neural-network)
- [Machine Learning](/en/glossary/machine-learning)
- [Transformer](/en/glossary/transformer)
- [Large Language Model](/en/glossary/large-language-model)
- [Computer Vision](/en/glossary/computer-vision)

[Back to the AI Glossary](/en/glossary)