Deep Learning

Deep learning is a specialised branch of machine learning that has driven the most dramatic advances in artificial intelligence over the past decade. It uses neural networks with many successive layers, each one learning to recognise progressively more complex features in the data. That layered structure is where the word "deep" comes from: it refers to the depth of the network, meaning the number of layers, rather than any particular sophistication of thought.

Before deep learning became practical, many AI tasks such as accurately recognising speech, translating languages, or identifying objects in photographs were considered extremely difficult. Deep learning changed that. Given enough data and computing power, deep neural networks can learn representations of the world that are rich enough to handle these tasks at or above human-level accuracy. The technology underpins voice assistants, large language models, image generators, and many medical diagnostic tools.

How it works#

A deep learning model is built from layers of interconnected nodes. Raw data enters the first layer: pixel values for an image, word tokens for text, or waveform data for audio. Each layer applies mathematical transformations to the data and passes the result to the next layer. Through a training process called backpropagation, the network adjusts the strength of every connection based on how wrong its predictions were, repeating this process millions of times until its outputs become reliably accurate.

A key advantage over older machine learning approaches is that deep learning can learn its own features. A traditional system for recognising handwriting might need an engineer to manually define features like stroke width or loop shapes. A deep learning model discovers useful features on its own, often finding patterns that no human would have thought to look for. The transformer architecture, a particular kind of deep neural network, is what powers most modern large language models and image generators.

Examples#

Medical imaging. Several UK hospitals are piloting deep learning tools that analyse X-rays and scans. A model trained on hundreds of thousands of labelled medical images can flag potential tumours or fractures for a radiologist to review, helping to prioritise urgent cases in a busy department.

Voice recognition. When you speak to a smartphone assistant or a voice-activated device, a deep learning model is converting your spoken words into text in real time. The network processes the audio waveform through many layers, identifying phonemes, then words, then the overall meaning of what you said. This capability was largely impractical before deep learning became mainstream.

Why it matters#

Deep learning is the reason that the current generation of AI feels qualitatively different from what came before. Tasks that required painstaking hand-coding, or that simply could not be done reliably at all, are now handled accurately and at scale. For anyone working with AI tools today, whether as a user or someone evaluating them for a team, understanding that deep learning models learn from data rather than rules helps explain both their impressive capabilities and their known weaknesses, such as hallucination and sensitivity to unusual inputs.

Our tutorial How LLMs Work explains how deep learning is applied in large language models, and AI vs ML vs Deep Learning sets out how these ideas relate to one another.