What is deep learning?

Definition

What is deep learning?

Deep learning is a subfield of machine learning that processes data with multi-layer neural networks and recognises complex patterns on its own. Today it forms the basis for most breakthroughs in artificial intelligence.

The term “deep” refers to the depth of these networks – that is, the number of hidden layers between input and output. Classic neural networks use a few layers; deep learning models use hundreds to thousands.

How it works

How does deep learning work?

Deep learning processes data in layers: each layer recognises a more abstract pattern than the previous one. At the end there is a prediction, a classification or generated content.

Layers

A deep learning model consists of an artificial neural network with an input layer, many hidden layers and an output layer. The input passes through each layer, which transforms it into a more abstract representation.

In an image recognition model, the first layer recognises edges, the second layer shapes, the third layer objects. Each level builds on the features of the previous one.

Automatic feature recognition

Classic machine learning methods require developers to define the relevant features manually. Deep learning takes over this step itself and extracts the features directly from the raw data.

As a result, deep learning processes unstructured data such as images, audio or text without manual pre-processing – a decisive advantage over classic algorithms.

Backpropagation

During training, the model adjusts its weights via a method called backpropagation. The algorithm measures the error of the prediction and distributes it backwards through all layers.

With each pass the model corrects its weights until the predictions are precise enough. This learning process consumes millions of examples and often weeks of computing time.

Architectures

What deep learning architectures are there?

Deep learning comprises several architectures for different data types: CNNs for images, RNNs and LSTMs for sequences, transformers for language. Each architecture suits particular tasks.

Convolutional neural networks (CNN)

CNNs process images by recognising local patterns such as edges and textures via filters. They dominate computer vision and are found in facial recognition, image classification and medical image analysis.

Recurrent neural networks (RNN) and LSTM

RNNs process sequences such as text or audio and store information from previous time steps. Long short-term memory (LSTM) extends RNNs with a mechanism that retains important information over longer periods.

RNNs and LSTMs shaped language processing for many years – until transformers replaced them in most tasks.

Transformers

Transformers process sequences in parallel rather than step by step and use a mechanism called self-attention that captures relationships between all words in a sentence at the same time. Google researchers introduced the architecture in 2017 in the paper “Attention is all you need”.

Every modern large language model builds on transformers – from GPT and Claude to Gemini.

Distinction

What distinguishes deep learning from machine learning?

Deep learning is a subset of machine learning that uses neural networks with many layers. Classic machine learning works with simpler algorithms and requires manual feature selection.

Feature recognition

In classic machine learning, developers select the relevant features themselves – such as colours or edges for an image analysis. Deep learning skips this step and learns the features directly from the raw data.

This difference makes deep learning particularly strong with unstructured data such as images, audio and text, where manual feature definition would be practically impossible.

Data volume and computing power

Deep learning requires millions to billions of training examples and specialised GPU hardware. Classic ML often manages with a few thousand examples and a normal CPU.

That feeds directly into the costs: a classic ML model trains in minutes; a large deep learning model consumes weeks of computing time and millions of euros.

When which method pays off

With small datasets or clearly structured tables, classic methods such as random forests or support vector machines often deliver equally precise results – with considerably less effort.

Deep learning only develops its strengths with unstructured data and large training volumes. For most business intelligence tasks, classic ML is completely sufficient.

Language models

Why is deep learning behind large language models?

Large language models such as GPT, Claude or Gemini are transformer-based deep learning models. Without deep neural networks, machines could neither understand nor generate language.

Transformer architecture

The transformer architecture stacks dozens to hundreds of layers on top of each other, in which each layer re-weights word relationships. Only this depth allows a model to capture grammar, meaning and context at the same time.

The self-attention mechanism checks in every layer which words in the sentence count for the next token. Classic machine learning does not master this kind of processing.

Scaling via parameters

An LLM has hundreds of billions of parameters – each of them a weight in a deep neural network. The more parameters, the more nuanced the language capabilities.

Research results show a clear scaling effect: larger models solve tasks at which smaller ones fail. Deep learning is what makes this scaling technically possible in the first place.

Training on billions of texts

During training, an LLM learns which token is most likely to follow the previous sequence. To do so it processes terabytes of text from books, scientific papers and the open web.

Training runs on clusters with tens of thousands of GPUs over weeks or months. Only deep learning methods can capture patterns in this amount of data at all.

Training

How do developers train a deep learning model?

Developers train a deep learning model by showing it large amounts of data and gradually adjusting its weights until the predictions are correct. The process runs on specialised hardware.

Data

Deep learning models require millions to billions of examples. Image classifiers often make do with hundreds of thousands of images; language models process many terabytes of text from the internet.

The quality of the data largely decides the result. Poorly labelled data or distorted samples produce poor models – regardless of the architecture.

GPU and TPU

Training runs on GPUs (graphics processing units) or on TPUs (tensor processing units from Google), which carry out thousands of matrix operations in parallel. A standard CPU would simply be too slow.

Modern language models train on clusters with tens of thousands of Nvidia H100 GPUs. The training costs for a large LLM often exceed tens of millions.

Overfitting

If a model memorises the training data too precisely, it fails on new data – this problem is called overfitting. Techniques such as dropout, regularisation and validation sets keep the model generalisable.

Practice

Where do companies use deep learning?

Companies use deep learning wherever machines process images, speech or complex patterns – from quality control in production to medical diagnostics.

Image processing

Car manufacturers use deep learning for lane detection and pedestrian detection in driver assistance systems. Cameras on production lines detect defective components in real time.

Radiologists use deep learning models to identify tumours and fractures in X-ray and MRI images.

Speech and text

Voice assistants such as Siri and Alexa convert audio into text and understand commands via deep learning models. Translation services such as DeepL are also built on neural networks.

Customer support chatbots, automatic summaries and sentiment analysis run on transformer-based models today.

Industry and medicine

Google DeepMind uses AlphaFold to predict protein structures – a breakthrough for drug research. Banks detect fraud patterns in millions of transactions per second.

Energy suppliers forecast electricity consumption and optimise grid stability via deep learning models that evaluate weather and consumption data simultaneously.

Limits

What limits does deep learning have?

Deep learning requires huge amounts of data, enormous computing power and in many cases remains a black box. These three factors limit where its use pays off.

Data hunger

Without sufficient training data, a deep learning model does not learn reliable patterns. For rare diseases or niche applications, the data is often simply missing.

Transfer learning mitigates the problem by having developers adapt pre-trained models to smaller datasets.

Black-box problem

A deep learning model makes predictions without humans being able to trace in detail why. In regulated industries such as medicine or finance, that remains a serious problem.

The discipline of explainable AI (XAI) develops methods that make the decisions of deep networks at least partially transparent.

Energy consumption

Training a large language model consumes as much electricity as several households in a year. The carbon footprint of deep learning is increasingly criticised.

Development

How has deep learning developed since 2012?

Deep learning has gone through three big leaps since 2012: AlexNet proved the power of deep networks, the transformer revolutionised language, generative AI brought deep learning into everyday life.

AlexNet 2012

AlexNet won the ImageNet competition in 2012 by a clear margin and ushered in the era of modern deep learning. The network was the first to use GPUs to train a deep CNN.

Transformer 2017

Google researchers introduced the transformer in 2017 in the paper “Attention is all you need”. The new architecture displaced RNNs in almost all language tasks and paved the way for modern LLMs.

Generative AI since 2022

OpenAI released ChatGPT in 2022 and brought deep learning into everyday life. Since then, new models that generate images, videos, music and code have appeared every month – all based on deep learning.

FAQ

Frequently asked questions about deep learning