Deep learning has driven artificial intelligence forward considerably in recent years. Image recognition, speech recognition, machine translation, and large language models are essentially based on this approach. The "deep" refers to the depth of the neural networks used, which consist of many layers connected in sequence.
The breakthrough came in the early 2010s, when large amounts of data and powerful graphics processors became available. Since then, deep learning has prevailed over older methods in many areas.
How does deep learning work?
In classic machine learning, experts often define themselves which features a model should consider, for example certain key figures or color values in an image. Deep learning learns such features on its own from the raw data. The first layers of a network recognize simple structures, and later layers combine them into increasingly abstract patterns – from edges and shapes to entire objects.
This requires large amounts of data and a lot of computing power. Training usually takes place on specialized hardware such as graphics processing units (GPUs). Important architectures are convolutional neural networks for images and transformers, which have been known since 2017 and form the basis of modern large language models.
Typical examples
- Visual quality control in production, such as detecting scratches or defects
- Text recognition in scanned documents and receipts
- Converting spoken language into text in meetings or dictations
- Translating documents and websites
- Text assistants that create drafts or summarize content
What companies should keep in mind
- Check ready-made models: Training your own deep learning model from scratch is costly. Pretrained models that are adapted as needed are often sufficient.
- Explainability: Deep networks are considered hard to understand. Where decisions must be justified, Explainable AI methods help.
- Rights to data: Before training your own model, clarify whether the available data may be used for it, especially if it is personal data.
- Operating costs: Computing power for training and operation causes ongoing costs, whether in the cloud or on your own hardware.
How it differs
Deep learning is not the opposite of machine learning but a part of it. For well-structured tabular data, for example from accounting or inventory management, classic machine learning methods often deliver equivalent results with less effort and better traceability. Deep learning plays to its strengths above all with unstructured data such as images, audio, and text.
View AI training