GlossaryFloor 1 · The Modela solid block on its own: the prediction machineFloor 1 · The Model
neural network
No. 048 · v2026-08FR: réseau de neuronesA neural network is a long sequence of multiplications and additions, organised in layers, whose coefficients are adjusted until the outputs match the examples. Like a mixing desk with billions of faders, set automatically by ear.
What it is not
A neural network does not reproduce the brain. The name comes from an analogy of the nineteen-forties, very loose and long since outdated: there is no cell, no chemistry, no sleep, no forgetting, and the architecture that dominates today has no known biological equivalent. Nor is it a magic box: every operation is elementary and perfectly defined. What is opaque is not the computation, it is the reason why a given combination of coefficients produces a given behaviour.
In depth
What a neuron does
An artificial neuron does three things: it multiplies each of its inputs by a coefficient, it adds everything up, then it passes the result through a small function that breaks linearity. It is this last step that counts: without it, stacking a thousand layers would amount to exactly one, and the network could only represent proportional relations. The coefficients are called parameters, and they, and they alone, constitute what a model has learned.
Backpropagation
The adjustment is done by backpropagation. An example is presented, the output is compared to the expected answer, the gap is measured, then the network is walked backwards to work out the share of each coefficient in that gap, and the coefficient is corrected slightly in the right direction. Repeated billions of times on billions of examples, this very simple operation is enough to bring out behaviours nobody wrote. It is also what makes training expensive and use cheap: setting the mixing desk takes months, using it afterwards takes a fraction of a second.
Opacity misnamed
Opacity is therefore badly named. Nothing is hidden: every parameter can be read and every operation replayed. What is missing is a reading at the level where one would like to judge, that of reasons. You cannot point to a place in the network and say “this is where it decided to refuse the application”, because the information is not stored in one place but distributed over billions of values. It is a structural property, not an engineering shortcoming, and it is what makes it indispensable to measure behaviours rather than hope to inspect them.
Relations where the neighbours live
Check 3 questions · click your answer
Level 1 · Recognise
What exactly does what a network “has learned” contain?
Level 2 · Distinguish
Why is a neural network said to be difficult to explain?
Level 2 · Distinguish
Why does training a model cost incomparably more than using it?
Lexigraph, "Neural network", v2026-08, https://www.lexigraph.org/en/neural-network/, CC BY 4.0.