AI Glossary / AI and machine learning / Beginner

Parameter

Definition

A parameter is one of the internal settings a model learns from data during training: the dial positions, like the weights on inputs, that together encode everything the model knows.

What it means

A model is a machine with a fixed design and adjustable insides. The design, chosen by people, says what kind of machine it is; the parameters, learned from data, say what this particular trained copy of it concluded. A straight-line model has two parameters, slope and intercept; a large language model has hundreds of billions.

Training is the systematic turning of these dials: the model makes a prediction, measures its error, and nudges every parameter slightly in the direction that would have reduced the error, millions of times, until the dials rest at values that fit the data well.

Parameters are why models are portable and opaque at once. The trained model is just its parameter list, a file of numbers that can be copied and shipped anywhere, and yet no human can read what those numbers mean directly, which is why explanation tools exist.

For a business, parameters are the asset: the dataset shaped them, the compute bill paid for them, and the file holding them is the thing licensed, stolen, versioned and audited.

How it works

Choose a model design, which fixes how many parameters exist and what each one influences.

Initialize the parameters, usually to small random values, so training starts from ignorance rather than a guess.

Train: repeatedly predict, measure error, and adjust every parameter by a small step in the error-reducing direction.

Freeze the parameters at the end of training; a deployed model is a fixed set of dial positions applied to new inputs.

Version the parameter file like any critical asset, because a silent change in numbers is a silent change in behavior.

Three practical examples

• A two-parameter straight line predicts delivery time from distance; the learned slope is the minutes per mile the data believed.
• A churn model's parameter on support-ticket count is large and positive, quantifying how strongly complaints drive departure, all else equal.
• A foundation model ships as a parameter file of hundreds of gigabytes, and running it means applying those numbers to each request.

Think of it this way

Think of a bakery's recipe card for sourdough: the card's design says flour, water, salt, starter; the learned parameters are the grams and minutes this bakery settled on after a thousand loaves. Anyone can copy the numbers; reading them does not tell you the taste.

Worked example

the two-knob model

Suppose a firm predicts monthly delivery cost from parcels shipped, using a straight line. The figures are illustrative.

Training settles on two parameters: an intercept of eight hundred, the monthly fixed cost the data implies, and a slope of two point four, the cost per parcel.

Auditors can check these two numbers against the finance ledger, and the model earns trust precisely because its parameters are few enough to read.

The general lesson: when parameters are few, they are explanations; when they are billions, they are an asset to govern, and the governance questions are the same.

Illustrative case study

A manufacturing group discovered during an acquisition review that its prized predictive-maintenance advantage was, concretely, eleven parameter files and the training runs that produced them. The discovery changed the deal work: data-processing agreements were audited as the origin of the parameters, version histories proved which file was in production, and one division could not actually locate its current parameter file, which became an embarrassing but cheap fix. The group now treats trained parameters as intellectual property with the same registry, backup and access controls as its patents, because that is what they are.

Common misconceptions and mistakes

• Parameters are set by the engineers. The design and its hyper-choices are set by people; parameters are learned from data by training.
• More parameters always mean a better model. More parameters mean more capacity to learn and more capacity to memorize; data and discipline decide which wins.
• You can read what a big model knows from its parameters. At scale they are legible only through the model's behavior, which is why testing and explanation tools exist.

Business use

Treat trained parameter files as governed assets: versioned, backed up, access-controlled and linked to the data and code that produced them.

Ask for parameter provenance in vendor contracts: what data shaped these numbers, and what rights do you have in them.

Prefer readable low-parameter models where decisions must be explained to regulators or customers, and pay for opacity only where it earns.

Record parameter counts in project reviews as a proxy for compute cost, memorization risk and hardware needs.

People also ask

What exactly is a parameter?

One learned internal setting of a model, like a weight on an input; the full parameter set is the trained model itself.

Who chooses the parameters?

Training does, by adjusting them to reduce errors; people choose the model design, which fixes how many exist.

Is the parameter file the model?

Functionally yes: the design plus the parameter file reproduces the model's behavior exactly.

Why do parameter counts make headlines?

They proxy capacity, compute cost and memorization risk, though quality depends on data and training as much as count.

Where do hyperparameters fit?

They are the settings people choose before training, like the learning rate; parameters are what training then learns under those choices.

Check your understanding

What is a parameter, and why does a business own its parameter files like intellectual property?

Answer: A parameter is one internal setting learned from data during training, and the full set of them is the trained model in portable form. The files deserve IP-grade governance because they are the distilled product of the company's data and compute spend: they are what gets licensed, stolen, versioned and audited, and a silent change in them is a silent change in every decision the model touches.

Related terms

• Parameter update
• Optimizer
• Training
• Weight
• Model
• hyperparameter
• weights

Learn next

Be Productive Using AI helps you ask what shaped a model before trusting what it says: https://aiforbusinessandlife.com/academy/be-productive-using-ai [C1]

• Model

More in AI and machine learning

• Overfitting - Overfitting is the failure where a model learns its training data too well, memorizing its noise and accidents along with its real patterns, so it performs...
• Packed data - Packed data is a storage layout where values are stored back to back in a fixed binary form, with no per-value overhead, so a million numbers occupy exactly...
• Pair Confusion Matrix - A Pair Confusion Matrix judges a clustering by looking at pairs of items: for every pair, did the clustering correctly put same-kind items together and...
• Parity learning - Parity learning is the task of learning which input bits are combined by an exclusive-or rule to produce a binary output.
• Partial least squares path modeling - Partial least squares path modeling is a structural-modeling approach that estimates relationships among constructs using weighted combinations of observed...
• Passive-aggressive learning - Passive-aggressive learning is a family of online learning methods that update a predictor as labeled examples arrive.

Sources

[S12] Aalto University, Dictionary of Machine Learning; [S2] Google, Machine Learning Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Parallelogram model Next: Parameter-efficient fine-tuning