What is a parameter in AI?
In one line
Ages 8 to 10: Parameters are a model's dials, and training turns each dial a little until the answers come out right more often.
Ages 11 to 13: In a weighted sum, each weight is a parameter, and training changes the weights to reduce the model's mistakes.
Ages 14 to 18: Parameters such as weights and biases are learned from data by optimization, while hyperparameters are settings people choose, usually by testing them on a validation set.
A parameter in AI is a number inside a model that the model learns during training. Weights are the most common kind: each one sets how much an input counts. Training nudges the parameters, little by little, until the model's answers fit its examples better. Settings people choose before training are called hyperparameters.
Last reviewed: 20 October 2026 6 min read
The idea in one picture
Think of a sound mixer with a row of dials. Each dial sets how loud one instrument is. A model's parameters are its dials, except that nobody turns them by hand: training turns them.
A tiny fruit-scoring model with three parameters:
score = 2 × sweetness + 1 × crunch − 3
The 2 and the 1 are weights; the −3 is a bias, a number that shifts every score. For an apple with sweetness 4 and crunch 5: 2 × 4 + 1 × 5 − 3 = 8 + 5 − 3 = 10.
If the training examples show crunch matters more than the model thinks, training can raise that weight from 1 to 2. The same apple then scores 2 × 4 + 2 × 5 − 3 = 15. Change a parameter and every answer moves.
How it works, step by step
- Start with rough or random parameter values.
- Predict an answer for a training example.
- Measure the error with a loss, a number for how wrong the answer was.
- Nudge each parameter slightly in the direction that makes the loss smaller.
- Repeat over many examples and many rounds.
- Freeze and use. After training the parameters stay fixed while the model makes predictions.
A little deeper (ages 11 to 13)
Parameters add up fast. A simple model that sorts small 32 × 32 color pictures into 10 categories looks at 3,072 numbers per picture. It needs one weight for every number and category, 10 × 3,072 = 30,720, plus one bias per category: 30,730 parameters. Neural networks with many layers have far more.
A parameter count says how big a model is, not how good it is. More parameters let a model capture more detail, but they also need more examples, or the model can overfit.
Going further (ages 14 to 18)
Parameters are learned from data. Hyperparameters are chosen by people: for example, the number of neighbors k in a nearest-neighbor classifier, or how strongly to penalize large weights. Hyperparameters are compared on a validation set, not the test set, so that the final test still measures performance on new data.
Learning works by gradient descent. Backpropagation calculates, for every weight, which way a small change would move the loss, and each weight steps the other way. Repeated many times over many examples, those small steps set every parameter.
Watch the vocabulary. In programming, a parameter is a named input in a function's definition, which is a different idea. And one widely used machine-learning library uses "parameter" for settings that statisticians would call hyperparameters.
Common mix-ups
- Mix-up: "Parameters and hyperparameters are the same."Fact: Parameters are learned; hyperparameters are chosen before training.
- Mix-up: "More parameters means a smarter model."Fact: More parameters mean more capacity, which needs more data and testing.
- Mix-up: "Parameters are like facts the model knows."Fact: They are numbers that shape how inputs become outputs.
- Mix-up: "A parameter in code is the same thing."Fact: In programming it names a function's input.
Try it (no device, 5 minutes)
Play "find the dial". One person picks a secret number, the weight, such as 3. The other calls out inputs, and the first replies with input × weight: "4?" "12." "7?" "21." After each answer, the guesser adjusts their guess for the weight. How many tries until you find it? For a harder round, add a secret bias: input × weight + bias.
Quick check
Using score = 2 × sweetness + 1 × crunch − 3, what does a pear with sweetness 5 and crunch 2 score?
Answer: 2 × 5 + 1 × 2 − 3 = 10 + 2 − 3 = 9.
2 × 5 + 1 × 2 − 3 = 10 + 2 − 3 = 9.In that rule, how many parameters are there, and what are they?
Answer: Three: the weights 2 and 1, and the bias −3.
Three: the weights 2 and 1, and the bias −3.Is the number of layers in a neural network a parameter or a hyperparameter?
Answer: A hyperparameter. People choose it before training; training does not learn it.
A hyperparameter. People choose it before training; training does not learn it.Questions
A parameter is a number inside an AI model that is learned during training. Weights are the most common kind: they set how much each input counts toward the answer.
Parameters, such as weights and biases, are learned from the training data. Hyperparameters are settings people choose before training, such as how many neighbors to compare or how strongly to penalize large weights.
It is how many learned numbers the model has, which is a measure of its size. A bigger count lets a model capture more detail but needs more data and more computing, and it does not by itself make the model better.
The model makes predictions, a loss measures how wrong they are, and each parameter is nudged a little in the direction that reduces the loss. Repeating this over many examples sets all the parameters.
Both are parameters. A weight multiplies an input to set how much it counts, and a bias is an extra number added to shift the result.
Mechna Academy is a city of machines for curious children aged 8 to 18, opening first to ages 8 to 10. In each mission, a child guesses how a machine decides, then opens its Peek Hatch to check, changes one thing and tests it on a case it has never seen. Children learn how AI works this way, and they never chat with an AI.
For grown-ups: join the waitlistPlaces are limited. The waitlist is for parents and guardians 18 and older.