Learn AI /Neural Networks /Weight
Weight
- Intermediate
- Neural Networks
- Technical readers
- Policy officers
Also called: weights, model weight, model weights.
Back to the milestone you came from
In five seconds
What it is
A weight is a learned number that scales how strongly one input signal affects the next calculation in a neural network.
In plain words
A weight is a number that tells the model how much one incoming signal should count in the next step.
An everyday example
A traveller in Dakar asks three friends how to get across the city. One friend cares most about speed, another strongly avoids tolls, and a third pays close attention to safety at night. The traveller combines all three suggestions, but gives different importance to each one depending on the journey. On a workday morning, speed may count most; late at night, safety may count more.
A neural network does not literally hear advice or consciously choose priorities for a trip. Also, in a model the weights are fixed numbers at inference time and are applied automatically to many signals at once, not reconsidered in a human-style discussion.
Remember thisA weight is a learned number that scales an input's contribution. Positive, negative and near-zero weights can strengthen, oppose or largely suppress that signal in the next calculation.
Understand it
Understand it
Why it matters
If you procure, govern, audit or use AI systems, weights matter because they are part of how a model turns inputs into outputs. Changes to weights can change performance, errors, bias patterns, energy use after pruning, and whether two deployments of 'the same model' actually behave the same way.
How weights change the strength of signals
After viewing, the reader should be able to say that a weight multiplies an input, several weighted inputs are combined, and the result is passed onward.
01
Inputs arrive
Several numbers enter a calculation from earlier parts of the network.
Render three source boxes on the left with numeric values and arrows pointing to the centre.
02
Each input is scaled
Each arrow has its own weight, which multiplies the input value.
Place a weight label on each arrow and show the multiplication near the centre.
03
Values are combined
The scaled values are added together to make one intermediate result.
Show the three products converging into one central circle with the sum displayed.
04
Result moves forward
That combined result is then passed to the next step, often after adding a bias and applying an activation function.
Draw an arrow from the centre to a right-side output box, with a small note below saying 'bias and activation may follow'.
Show a left-to-right flow diagram with three input boxes labelled 'input A = 2', 'input B = 5', and 'input C = 1'. Each connects by an arrow to a central calculation node. Place a small label on each arrow: 'weight 0.8', 'weight -0.4', and 'weight 1.2'. Next to the central node, show the multiplied values '2×0.8=1.6', '5×-0.4=-2.0', and '1×1.2=1.2', then a line reading 'sum = 0.8'. From there, draw an arrow to an output box labelled 'signal sent onward'. Use colour only as a secondary cue; also vary line thickness so larger absolute weights look stronger.
On hover or tap, each input arrow highlights and the matching multiplication appears bold so the reader can see how one weight changes one contribution.
Describe this without the diagram
The diagram shows three inputs flowing into one calculation. Input A has value 2 and weight 0.8, so it contributes 1.6. Input B has value 5 and weight minus 0.4, so it contributes minus 2.0. Input C has value 1 and weight 1.2, so it contributes 1.2. These contributions are added to produce 0.8, which is passed onward in the network. The key point is that a weight changes both the strength and sometimes the direction of an input’s effect.
With reduced motion: A static left-to-right diagram shows three inputs, their weights, the multiplied contributions, the summed result, and the outgoing signal, without any highlighting or animation.
In a neural network, numbers come in as inputs. Each connection into a calculation can have its own weight. That weight multiplies the incoming value, so some signals count more and some count less.
A single weight is one such number. “Weights” usually means the whole collection of these learned numbers across a model. “Model weight” can mean one specific learned number, but in practice “model weights” often means the saved file or files containing many learned numbers for a trained model. Those uses are related, but not identical.
During a forward pass, the network combines inputs with weights, often adds a bias, and then applies an activation function, which is a rule that transforms the result. What comes out of one layer becomes input to the next. The final output might be a score, a class label, or text probabilities.
Weights are adjusted during training so that the model’s outputs better match examples it has seen. After training, the weights store much of what the model has statistically fitted from data. That does not mean they store clear human-readable rules, and it does not guarantee the output is correct.
How “Choosing a route across a city” maps onto the real thing
| In the story | In the AI system |
|---|---|
| how much the traveller listens to each friend | the weight on each input signal |
| the friends' route suggestions | incoming input values or features |
| combining several suggestions into one route choice | computing a weighted sum before the next step |
| changing priorities for different situations | different learned weights in different parts of a network |
A worked example
For the lessons-learned note, the duty officer checks a simplified neuron that scores whether a mock enquiry concerns delivery disruption. This illustration omits the bias and activation function so that the effect of the weights remains visible.
Starting from: delivery-term score = 3, location-match score = 2, unrelated-topic score = 4
- The learned weights are 1.5 for the delivery-term score, 0.7 for the location match and -1.0 for the unrelated-topic score.
- Scale each input by its weight: 3 × 1.5 = 4.5, 2 × 0.7 = 1.4 and 4 × -1.0 = -4.0.
- Combine the weighted contributions: 4.5 + 1.4 - 4.0 = 1.9.
- The value 1.9 is passed to the next operation. Its eventual effect depends on any bias, activation function and later layers in the model.
Result: The positive weights strengthen the delivery-related signals, while the negative weight makes unrelated subject matter reduce the combined value. The weights control each input's contribution, rather than independently deciding the final classification.
In a United Nations workplace
During a lessons-learned exercise after a pilot, a duty officer in a procurement unit drafts a note for the file before close of business. The pilot used an approved model to route routine supplier enquiries. Inside the model, learned weights determine how strongly signals such as delivery terminology, invoice language and unrelated subject matter affect each routing calculation.
When the vendor says that retraining changed the weights, the officer should understand that the model's internal numerical priorities changed. A signal may therefore contribute more, less or in the opposite direction, altering which enquiries are routed to each queue even if the workflow itself looks unchanged.
Before you try thisUse synthetic or properly cleared test material for the exercise. Do not enter bids, commercial details, personal data or other procurement-sensitive material into a tool not approved for that information. Doing so can compromise confidentiality, fairness and the integrity of the procurement process.
Where you may hear it
- A vendor demo during a procurement meeting
- A lessons-learned exercise after an artificial intelligence pilot
- An ICT governance board reviewing a proposed model update
- A technical briefing before revised model files are deployed
Avoid confusion
Avoid confusion
Often confused with
Bias (in a neural network)
A bias is an extra adjustable number added to a neuron's weighted input so the neuron's output can shift even when all inputs are zero.
Parameter
A parameter is a learned numerical setting inside a model that helps shape which token it is likely to produce next.
Feature
A feature is a piece of input information a model uses to help produce an output, such as a prediction, score or grouping.
Common misconceptions
Common misconception: A weight is the same as a feature.
Correction: A feature is an input signal or measured property, such as word presence or pixel intensity. A weight is the learned number that tells the model how strongly that feature or intermediate signal influences the next calculation.
Common misconception: Weights are the same as bias terms.
Correction: Both are learned parameters, but they play different roles. A weight scales an incoming signal, while a bias shifts the combined result up or down even when inputs are zero.
Common misconception: Model weights are human-readable knowledge statements.
Correction: Weights are usually just arrays of numbers. They encode statistical fitting in a distributed form, not neat sentences like a policy manual or rulebook.
Common misconception: If you have the model weights, you automatically know why every output happened.
Correction: The weights are necessary to reproduce the model, but large networks have many interacting weights. Inspecting them directly often does not give a simple explanation for one specific output.
Quick check
Which statement best distinguishes a single weight from model weights as a released artefact?
One weight is one learned numerical factor in a calculation. In practice, 'model weights' often means the stored tensors or files containing many learned parameters for the whole trained model. The other options confuse weights with features, bias terms, activation functions, or deployment stages.
Go deeper
Go deeper
The simple version
A weight is a number that tells the model how much one incoming signal should count in the next step.
The technical version
In a neural network, a weight is a learned parameter attached to a connection or operation that linearly scales an input or intermediate activation before combination with others, often followed by a bias term and a non-linear activation function.
Reviewed external resources
EncyclopaediaWeighting (opens in a new tab)Wikipedia
This link gives a broad background on weighting as a general idea and suits readers who want a basic reference before focusing on neural networks.
Intermediatelink checked 2026-07-31
InteractiveTensorFlow Playground (opens in a new tab)TensorFlow
This interactive tool lets readers change weights and see outputs update, which suits hands-on learners who want to experiment.
Starterlink checked 2026-07-31
VideoNeural networks, explained visually (opens in a new tab)3Blue1Brown
This video gives a visual intuition for how weights shape signals in a network and suits readers who learn best from diagrams and animation.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Weight (opens in a new tab)YouTube search
This is a YouTube search results page, so it helps readers compare different video explanations of weight and suits people looking for a teaching style that fits them.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Weight (opens in a new tab)Reddit search
This is a Reddit search results page, so it shows community questions and informal explanations about weight and suits readers who want to see common confusions and discussions.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
A weight is a learned number that scales an input's contribution. Positive, negative and near-zero weights can strengthen, oppose or largely suppress that signal in the next calculation.