Learn AI /Neural Networks /Forward pass
Forward pass
- Deeper
- Neural Networks
- Technical readers
Also called: feedforward.
Back to the milestone you came from
In five seconds
What it is
The calculation that carries input numbers through a neural network’s layers to produce an output such as a score, label or next-token distribution.
In plain words
A forward pass is the model’s one-way calculation from input numbers to output numbers.
An everyday example
Write two numbers on a card and pass it along a row of calculator stations. Each station applies its own rule, such as multiplying, adding or replacing a negative result with zero, then hands the new numbers to the next station. The last station turns them into two final scores. That trip from the original numbers to the scores is the forward pass. Changing the stations' rules would be a separate process.
A library routing process uses people applying explicit categories and may involve reasoning, clarification and external knowledge. A neural network forward pass is only a numerical computation with fixed parameters at that moment. It does not look up shelves or explain its route unless extra systems are added. Also, real networks process many values in parallel, not as a single request moving desk to desk.
Remember thisA forward pass carries input numbers through the network to produce an output using current parameters. It does not, by itself, update those parameters.
Understand it
Understand it
Why it matters
If you use, evaluate or procure AI systems, the forward pass is the core step that actually produces an output. Understanding it helps you ask practical questions: What inputs does the system use? What output does it return: a score, label or probability distribution? Does the model simply compute with fixed parameters here, or is it also being updated elsewhere? This matters for testing, latency, cost, explainability and risk controls. For example, if outputs change, you need to know whether the cause is new input data, different model parameters, or post-processing after the forward pass. It also helps you separate the model’s computation from training and from broader deployment workflows.
How numbers move through a forward pass
After viewing this diagram, a reader should be able to say that a forward pass takes input values, transforms them layer by layer using weights, bias and activation functions, and produces an output without changing the parameters.
01
Input enters
The process starts with input values such as features, pixels or token vectors.
Draw a left column of 3 to 5 circles labelled with example input values, such as x1=0.8, x2=0.1, x3=0.6. Use arrows pointing to the next column.
02
Weighted sums
Each unit in the next layer multiplies incoming values by weights and adds them together.
In the second column, show 2 to 3 circles. On at least one incoming arrow, label a weight such as w=1.2 or w=-0.4. Next to one target circle, show a compact formula like (x1·w1 + x2·w2 + x3·w3).
03
Bias and activation
A bias shifts the sum, then an activation function transforms it so the network can model non-linear patterns.
Beside each circle in the hidden layer, add a small badge '+ b' followed by an activation label such as ReLU. Show a sample transition like 0.3 -> ReLU -> 0.3 and -0.2 -> ReLU -> 0.
04
Repeat through layers
The output values from one layer become the input values for the next layer.
Add a third column with the same pattern and arrows from hidden layer 1 to hidden layer 2. Use consistent spacing and directional arrows to emphasise one-way flow.
05
Produce final output
The last layer returns scores, a value, or a distribution over classes. A softmax step may convert scores into probabilities.
Draw a final output column with 2 to 3 labelled outputs such as 'urgent 1.8', 'routine 0.6', then an optional softmax box turning them into 'urgent 0.77', 'routine 0.23'. Add a note: 'parameters unchanged during this pass'.
A left-to-right flow diagram with columns for input, hidden layer 1, hidden layer 2 and output. Arrows carry example numbers forward. Each middle column shows three mini-actions: weighted sum, add bias, apply activation. The final column shows output scores and an optional softmax step. Read from left to right.
Hovering or tapping a layer highlights only the arrows and formulas used at that stage, and a toggle shows outputs either as raw scores or after softmax.
Describe this without the diagram
This diagram explains a forward pass as a one-way flow of numbers through a neural network. First, input values appear in a left column. Next, each later layer performs three operations in order: multiply inputs by weights and add them, add a bias term, and apply an activation function. The resulting values move to the next layer and the same pattern repeats. In the final layer, the network outputs scores or values. An optional softmax step converts class scores into a probability distribution. The key message is that the model computes outputs using existing parameters; it does not update those parameters during this process.
With reduced motion: Show the same left-to-right diagram as a static image with all arrows visible at once, fixed labels for weighted sum, bias and activation in each layer, and a final note stating that the pass computes outputs without changing weights.
A forward pass is the part of a neural network that turns input into output. The input might be pixel values from an image, numbers from a sensor, or token representations from text. Those values enter the first layer, pass through one or more middle layers, and produce final output values.
In a typical dense layer, the network combines numbers using weights and a bias. A weight scales an input’s contribution and may change its sign. A bias shifts the weighted sum up or down. The layer then usually applies an activation function, often a non-linear rule that helps the network represent more complex patterns.
This process repeats layer by layer. The output of one layer becomes the input to the next. In a classifier, the final numbers may be class scores, and a function such as softmax may convert them into values that can be interpreted as a probability distribution over classes. In other tasks, the final output could be a single number, a vector, or the next-token scores in a language model.
A forward pass does not update the model’s parameters. It only uses the current weights and biases to compute outputs. That is why it is different from backpropagation, which computes gradients used to adjust parameters during training, and from the broader term inference, which usually means using a trained model in practice and may involve one or many forward passes.
How “A row of calculator stations” maps onto the real thing
| In the story | In the AI system |
|---|---|
| the visitor’s note | the input values entering the network |
| each desk applying its own routing rule | each layer combining inputs with weights, bias and an activation function |
| the output from one desk becoming the input to the next desk | layer-by-layer propagation through the network |
| the final shelf location | the network’s final output or prediction |
A worked example
A tiny network helps queue public meeting notices for a delegation briefing. It classifies each notice as "include" or "hold" using two features: whether the agenda matches the delegation's brief and a recency score.
Starting from: x1 = 1 for an agenda match; x2 = 0.4 for moderate recency
- Hidden unit: (1 × 0.9) + (0.4 × 0.5) - 0.2 = 0.9. ReLU leaves the activation at 0.9.
- Include logit: (0.9 × 1.3) - 0.1 = 1.07.
- Hold logit: (0.9 × -0.8) + 0.2 = -0.52.
- Softmax converts the two logits into approximate probabilities: include 0.83 and hold 0.17.
Result: This forward pass gives the higher score to "include". Under the current parameters, the notice moves up the review queue, although the officer still decides whether it belongs in the briefing.
In a United Nations workplace
A national officer in a shared services centre is preparing a delegation briefing before capitals reply. An approved internal classifier takes numerical features from public meeting notices, passes them through its layers and produces scores for whether each item should be reviewed for inclusion or held aside.
Each notice triggers a forward pass. The model transforms the input numbers into category scores using its current weights and biases. It is producing an output, not learning from the notice or changing its parameters.
Before you try thisDraft positions, instructions from capitals and personal or confidential correspondence must not be entered into a tool unless it is approved for that material. The scores may order the review queue, but the national officer remains responsible for what goes into the briefing.
Where you may hear it
- A vendor demonstration during a procurement meeting
- An ICT governance board review of model architecture
- A debugging session before an internal classifier is uploaded to the portal
- A performance review focused on response time
Avoid confusion
Avoid confusion
Often confused with
Backpropagation
A method that propagates derivatives backwards through a model to compute how changes to each weight would affect the loss.
Inference
Using a trained model on new input to produce an output such as a score, category, translation or generated text.
Input, hidden and output layers
These layers organise how a neural network turns input numbers into intermediate calculations and then into an output such as a score, label or predicted value.
Common misconceptions
Common misconception: A forward pass means the model is learning from the input it just received.
Correction: Usually it does not. In a standard setup, the forward pass only computes outputs using the current weights and biases. Learning happens when training also includes a loss calculation and backpropagation, followed by an optimiser step that updates parameters.
Common misconception: Forward pass and inference are exactly the same thing.
Correction: They are related but not identical. A forward pass is the internal computation from input to output. Inference is the broader use of a trained model in practice and may include preprocessing, one or more forward passes, decoding, thresholds, post-processing and system-level checks.
Common misconception: The network produces probabilities automatically at every output.
Correction: Not always. Many models first produce raw scores, often called logits. These can be turned into a probability distribution with a function such as softmax, but some tasks use raw values directly, such as regression or ranking.
Common misconception: A forward pass always goes strictly through simple stacked layers only once.
Correction: That is true for many feedforward networks, but the idea also applies in more complex architectures. Convolutional, recurrent and transformer-based models still perform a forward computation, even though the internal structure may include recurrence, attention or branching.
Quick check
Which statement best describes a forward pass in a neural network?
A forward pass computes outputs from inputs using the model’s existing parameters. Weight updates belong to training steps after backpropagation. Outputs are not always probabilities unless an additional step such as softmax is used, and forward passes occur in many network types, not only very simple ones.
Go deeper
Go deeper
The simple version
A forward pass is the model’s one-way calculation from input numbers to output numbers.
The technical version
Given an input tensor, the model applies its parameterised functions in order: each layer computes a transformation such as an affine operation using weights and bias, followed by a non-linear activation or another layer-specific operation. Intermediate activations are passed to later layers until the network produces output logits, values or vectors. In classification, logits may then be passed through softmax to obtain a probability distribution over classes. In training, the forward pass also creates the intermediate values needed for loss computation and later gradient calculation, but the forward pass itself does not change parameters. In implementation, it may be run for one example or a batch, and its cost is often measured in latency, memory use and floating-point operations.
Reviewed external resources
EncyclopaediaForward pass (opens in a new tab)Wikipedia
This Wikipedia link gives a broad reference-style overview of the term and suits readers who want a quick definition plus related concepts.
Intermediatelink checked 2026-07-31
InteractiveTensorFlow Playground (opens in a new tab)TensorFlow
This TensorFlow Playground interactive lets readers change inputs, weights and layers to see how forward computation affects outputs, which suits hands-on learners.
Starterlink checked 2026-07-31
VideoNeural networks, explained visually (opens in a new tab)3Blue1Brown
This 3Blue1Brown video gives a visual intuition for how signals move through neural network layers and suits visual learners.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Forward pass (opens in a new tab)YouTube search
This YouTube search results page helps readers find multiple video explanations of forward pass at different difficulty levels, which suits people comparing teaching styles.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Forward pass (opens in a new tab)Reddit search
This Reddit search results page helps readers see common questions, confusions and informal explanations about forward pass, which suits readers who want community discussion.
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 forward pass carries input numbers through the network to produce an output using current parameters. It does not, by itself, update those parameters.