GlossaryFloor 1 · The Modela solid block on its own: the prediction machineFloor 1 · The Model
deep learning
No. 047 · v2026-08FR: deep learning (apprentissage profond)Deep learning is the branch of machine learning that stacks up many layers of computation, each one describing the data a notch more abstractly than the one before. Like a chain of reviewers where the first sees strokes, the next shapes, the last a face.
What it is not
Deep learning is not “machine learning, only better”: it is hungrier for data and for computation, harder to explain, and on tabular data it is regularly beaten by simpler methods. The word “deep” says nothing about the quality of the result nor about any depth of understanding: it describes a feature of architecture, the number of layers stacked, and nothing else.
In depth
What the layers bring
What the layers bring is the automatic construction of descriptors. In classic learning, a person decides which characteristics to measure: the surface area of a home, its number of rooms, its distance from a station. That work, long and decisive, is called feature engineering. A deep network does without it: each layer itself learns which description of the input is useful to the next layer. That is what unlocked images, sound and text, where nobody knew how to say by hand what had to be measured.
The price
The price is threefold, and it must be known before choosing this route. Far more examples are needed, because there are far more values to adjust. Specialised hardware is needed, graphics cards having precisely the architecture that suits these repetitive computations. And direct explanation is lost: a decision tree can be read, a network holding a few tens of billions of values cannot, which weighs heavily as soon as a decision has to be justified to the person it concerns.
The strategic consequence
The strategic consequence is easy to state and often ignored. On tables of structured data, which remain the most common format in business, classic methods are frequently more accurate, faster and more explainable. Deep learning imposes itself where the input is raw and of high dimension: an image, a recording, free text. Choosing depth by default, because that is the word going around, amounts to paying three times the price for an inferior result.
Relations where the neighbours live
- Requires
- Floor 1 · The Modela solid block on its own: the prediction machineneural networkFloor 1 · The Modela solid block on its own: the prediction machinetraining data
- Often confused with
- Floor 1 · The Modela solid block on its own: the prediction machinemachine learningFloor 1 · The Modela solid block on its own: the prediction machinegenerative AI
- Related comparisons
- Machine learning or Deep learning
Check 3 questions · click your answer
Level 1 · Recognise
What does the word “deep” designate in “deep learning”?
Level 2 · Distinguish
You have to predict a risk of non-payment from a table of thirty columns. Is deep learning called for?
Level 2 · Distinguish
What does depth make unnecessary, compared with classic learning?
Lexigraph, "Deep learning", v2026-08, https://www.lexigraph.org/en/deep-learning/, CC BY 4.0.