GlossaryFloor 1 · The Modela solid block on its own: the prediction machineFloor 1 · The Model
machine learning
No. 046 · v2026-08FR: machine learning (apprentissage automatique)Machine learning consists in no longer writing the rules of a program but giving it examples so that it derives them itself. Like learning to recognise an edible mushroom by seeing a thousand of them, rather than by reading their description.
What it is not
Machine learning is not deep learning, which is only one family within it: a regression on three spreadsheet columns is part of it, and many systems in production use nothing more. Nor is it learning in the human sense: nothing is understood or memorised, numerical values are adjusted to reduce a measured gap between what the system predicts and what the examples show. And it is not a machine that goes on learning as it runs: unless something is explicitly set up for that, a trained system is frozen, and it draws no lesson from what you give it today.
In depth
The reversal
The reversal is as follows. A classic program receives data and rules written by a person, then produces answers. A learned system receives data and known answers, and produces the rules. What you obtain is called a model. The practical consequence is immediate: behaviour is no longer corrected by changing code but by changing examples, and a good part of the work moves towards assembling and labelling data. It is also why these systems fail differently: instead of a reproducible error that can be read in the code, you get a drop in performance on one category of cases, which has to be measured before it can be seen.
Three families
Three families cover most uses. Supervised learning starts from labelled examples, where the right answer is known: this is the regime of email classification or of predicting an amount. Unsupervised learning looks for structures in unlabelled data, groups of customers who resemble one another for instance. Reinforcement learning, finally, learns by trial and reward, and serves in particular to adjust the behaviour of language models after their main training. These families are not opposed: a real system often combines several of them.
Two traps
Two traps come up constantly. The first is overfitting: a model that reproduces its training examples perfectly can collapse on new cases, because it retained their peculiarities rather than the rule. This is why data never seen is always set aside, in order to measure honestly. The second is drift: the world changes, the data arriving stops resembling that of training, and performance falls without any alert being raised. A learned system is therefore never delivered once and for all, it is watched.
Relations where the neighbours live
Check 3 questions · click your answer
Level 1 · Recognise
What most clearly distinguishes a learned system from a classic program?
Level 2 · Distinguish
A model gets excellent results on its training data and poor results in production. What is this problem called?
Level 2 · Distinguish
Your assistant gets a case wrong. Does correcting that case in the conversation improve the system for other users?
Lexigraph, "Machine learning", v2026-08, https://www.lexigraph.org/en/machine-learning/, CC BY 4.0.