Deep learning is the part of machine learning that works with neural networks built from many stacked layers. The “deep” refers to that depth and to nothing else: not cleverer, not more conscious, just more layers between input and result. Almost everything that makes the method powerful, and everything that makes it expensive, follows from that single structural property.
Until around 2012 deep learning was an academic niche. Then enough data and enough graphics cards came together, and the niche turned into the foundation of practically every AI application making headlines today.
What the depth buys you
Features are learned rather than built
In classic machine learning a human decides what the model should look at. An image is reduced to edges, colour distributions and shape measurements, a text to word counts. This handiwork is called feature engineering, and it consumes most of the domain expertise in a project.
A deep network takes that step over itself. It receives the raw data and works out along the way which features are useful. That is the real break. Not the higher accuracy, but the shift of effort from building features to collecting data.
Each layer builds on the one before
The structure is easiest to describe with images. Early layers respond to brightness jumps and edges. The next ones assemble corners and curves from those, then eyes and wheels, and right at the end faces or vehicles. None of these stages was prescribed by anyone, they emerge during training. That is exactly why the approach also works for speech and sound, where nobody could say which features would be the right ones.
More layers therefore mean the network can express relationships in more intermediate steps. On simple tasks that is wasted effort. On tangled relationships it is the difference between “works” and “does not work”.
What it costs
Appetite for data
A network expected to find its own features needs material to find them in. Where a classic method with well-built features gets by on a manageable dataset, a deep network demands considerably more examples. The usual way out is transfer learning: you take a network somebody else pre-trained on a huge collection and adapt only the last layers to your task. That brings the requirement into a range individual companies can reach.
Appetite for compute
Training means repeating a great many arithmetic operations a great many times. It runs on graphics cards or specialised accelerators, takes hours to weeks depending on size, and is rarely good on the first attempt. On top of that come the running costs: a large network needs compute for every single request, and that bill keeps arriving for as long as the application is live.
| Classic machine learning | Deep learning | |
|---|---|---|
| Features | built by hand | learned by the model |
| Data required | moderate | high, unless transfer learning |
| Hardware | ordinary servers | graphics cards or accelerators |
| Strong on | tables, key figures | images, sound, speech, text |
| Traceability | often good | difficult |
Where it wins and where it does not
It wins wherever the raw data is unstructured and nobody can name the relevant features up front.
- Images and video: recognising objects, spotting defects in production, sorting footage.
- Speech and sound: transcribing dictation, telling speakers apart, classifying noises.
- Text: capturing meaning instead of counting words. Everything that currently runs under generative AI belongs here.
It does not win on tabular business data. For revenues, master data and key figures in rows and columns, decision-tree methods regularly deliver results that are just as good or better, train in minutes rather than hours, and can be explained to a business department. The effort pays off just as poorly when few examples exist, when every decision has to be justified, or when the application must run on a small device with no meaningful compute.
The honest order in a project is therefore: a simple method first as the benchmark, then the question of whether the depth improves that figure enough to be worth its price.
The road to today’s language models
For a long time the problem with text was sequence. Older network architectures read word by word and lost the beginning of a long sentence on the way. In 2017 the transformer architecture arrived, which looks at a whole passage at once and weighs up for itself which words matter to each other. That made training parallelisable, and parallelisability was the precondition for everything that followed.
The second step was the realisation that the same architecture simply keeps getting better with more data and more compute instead of hitting a ceiling. Out of that came the large language models that now drive chatbots, assistants and translation. They are not a new method, they are deep learning at enormous scale, plus a two-stage training: general pre-training on vast amounts of text first, then fine-tuning towards helpful behaviour.
The downsides of depth travel along with it. A model that continues patterns does not distinguish between a supported statement and one that merely sounds plausible. That is the structural reason for the AI hallucination, and it is why serious applications hand the model its facts through retrieval augmented generation instead of relying on its memory.
Frequently asked questions about deep learning
Is deep learning the same thing as AI? No. Artificial intelligence is the umbrella term, machine learning is one part of it, and deep learning is a part of machine learning in turn. That marketing mixes the three words up does not change how they nest.
Do I need my own graphics cards? For training large networks yes, for most other things no. Adapting and running pre-trained models works through rented compute or ready-made services. Whether the data is allowed to leave the building for that belongs at the start of the project.
How many layers are enough? The task decides that, not ambition. An oversized network memorises the training data and fails on new cases. In practice you adopt a proven architecture for the data type at hand and change only what measurably helps.
Can I trace why the network decided that way? Only to a limited degree. There are methods that show which parts of the input tipped the balance, but they do not produce a justification in the human sense. Wherever a decision has to be justified, for instance under the EU AI Act, that is a hard selection criterion.
Conclusion
Deep learning has fundamentally changed how machines handle images, speech and text, and it underpins the models everyone is currently talking about. For tabular data in a mid-sized company it often remains the second-best choice, because simpler methods are faster, cheaper and more explainable there. The decision follows the data type, not the mood of the moment. Which method fits your data is exactly the question AI consulting for companies starts with. How we build software, and where we use AI while doing it, is described on our page about custom software development. What holds up in daily work is covered in the article AI in software development, and for anything beyond that there is a free consultation.