Learn AI /Neural Networks /Convolutional neural network
Convolutional neural networkCNN
- Deeper
- Neural Networks
- Technical readers
Also called: CNN, convnet.
Back to the milestone you came from
In five seconds
What it is
A layered model that reuses small filters across a grid-like input, combines local patterns into larger features, and produces task outputs from them.
In plain words
A CNN is a neural network that checks small nearby parts of an image or other grid, finds repeating patterns, and combines those findings into a final result.
An everyday example
A team in Nairobi receives thousands of paper survey forms. Instead of reading every page as one big block, one colleague places a small transparent window over the top-left part of each page and checks for simple marks such as a tick, a bold heading or a handwritten note. They slide the same window across the page, row by row, recording where each pattern appears. A second colleague looks at that pattern record to spot larger structures, such as a complaint section with several related marks close together. At the end, the team decides whether the form is about water access, school safety or something else.
Real CNN filters are learned numerical weights, not hand-designed visual checks by a person. The model does not literally recognise concepts the way a colleague reads a form. Also, modern CNNs usually process many filters and layers in parallel, and later layers can be hard for humans to interpret clearly.
Remember thisA CNN reuses learned filters across local parts of a grid, then combines the resulting feature maps into larger patterns and a task output.
Understand it
Understand it
Why it matters
You may hear about CNNs whenever an organisation uses AI on images, video frames, scans, medical images, satellite pictures or other grid-like data. For a non-specialist, the practical question is not the mathematics alone but whether this model type fits the task. CNNs can be efficient because they reuse the same learned filters across many locations, which often works well when the same pattern can appear anywhere in an image. That design choice affects data needs, compute cost, explainability limits, and failure modes. If you are choosing, procuring or reviewing a system, it helps to recognise that a CNN is built to detect local patterns and spatial structure, not to handle all data types equally well.
How a CNN turns pixels into a class prediction
After viewing this, a reader should be able to explain that a CNN reuses small filters across an input, builds feature maps, combines them in deeper layers, and then produces output scores.
01
Input grid
Start with a grid of numbers, such as pixel values in an image.
Draw a 6x6 square grid on the left with varied shading. Label it 'input pixels'.
02
Shared filter scans local areas
A small filter, for example 3x3, moves across the grid. At each location it calculates one new value from nearby inputs.
Overlay a coloured 3x3 square on one input region, then show the same square repeated at two other positions with arrows, labelled 'same weights reused'.
03
Feature map
The scan creates a new grid that becomes strong where that filter matched a pattern such as an edge.
To the right, draw a smaller 4x4 grid with some dark cells. Connect several scan positions to corresponding cells in this grid.
04
Deeper feature building
Multiple filters and layers produce several feature maps, then combine them into more useful patterns while often reducing spatial size.
Show three stacked feature maps feeding into two smaller stacked maps, with labels 'edges/textures' then 'parts/shapes'.
05
Output layer
Later layers convert the learned features into task outputs such as class scores, often normalised with softmax for classification.
Draw three output bars labelled with example classes such as 'road', 'building', 'tree', with one bar tallest.
Show a left-to-right pipeline: image grid, sliding filter over small patches, resulting feature maps, a smaller set of stronger combined features after repeated layers, then output scores. Read it as data moving through stages, with reused filters highlighted to show that the same detector is applied in many positions.
On hover or tap, highlight one filter and show all positions where the same weights are reused; a toggle switches between early-layer and deeper-layer feature maps.
Describe this without the diagram
The diagram is a left-to-right flow. First, a grid of pixel values represents an input image. Next, a small 3 by 3 filter is shown moving over different parts of that grid, with a note that the same weights are reused at each position. That produces a feature map, a new grid whose values are high where the filter matched a local pattern. Several feature maps then feed into deeper layers that combine simpler patterns into larger ones and often reduce spatial size. Finally, an output layer turns those features into class scores, with one class highest. The main lesson is that CNNs detect local patterns first and build toward larger features through repeated layers.
With reduced motion: A static five-panel diagram shows the same stages without animation: input grid, one highlighted 3 by 3 filter with three example positions, one resulting feature map, a stack of deeper feature maps becoming smaller, and final output bars.
A convolutional neural network is a kind of neural network built for data arranged like a grid. The most common example is an image, which is a grid of pixel values, but the idea can also apply to other structured data. Instead of looking at every input value separately at first, it examines small nearby groups.
The key operation is a convolution. In simple terms, a small set of numbers, often called a filter or kernel, slides across the input. At each position, it multiplies nearby input values by the filter values and adds them up. This produces a new grid called a feature map. Different filters can respond to different local patterns, such as edges, corners or textures.
After several layers, the model combines simple local patterns into larger and more abstract ones. Early layers may react to lines or colour changes. Later layers may react to shapes or arrangements of parts. Pooling or strided steps may reduce spatial size, retaining selected or aggregated signals while lowering computation. Then one or more later layers turn those signals into an output such as class scores.
What goes in is usually a grid of numbers, such as pixel intensities. What happens is repeated local scanning, weighted combination, non-linear activation, and gradual feature building across layers. What comes out depends on the task: a label, a probability-like score after softmax, a detected region, or another structured prediction. CNNs are especially useful when nearby input values matter and the same kind of pattern can appear in many positions.
How “Inspecting printed forms with a reusable window” maps onto the real thing
| In the story | In the AI system |
|---|---|
| small transparent window moved across the page | filter or kernel sliding across the input |
| checking for a simple mark at each position | computing a local weighted match during convolution |
| pattern record showing where marks appeared | feature map |
| second colleague combining simple marks into a larger section pattern | deeper layers combining low-level features into higher-level features |
A worked example
Classifying a 6 by 6 black-and-white symbol as either a vertical line or a horizontal line.
Starting from: A 6 by 6 grid in which every cell in the third column has value 1 and all other cells have value 0.
- Apply a 3 by 3 vertical-line filter whose rows are [-1, 2, -1]. Slide the same nine learned weights across every valid position in the input, producing a 4 by 4 feature map.
- When the centre of the filter aligns with the bright column, its response is 6. Positions immediately beside the column produce negative responses, while positions farther away produce 0.
- Apply a rectified linear activation. This changes negative responses to 0 while preserving the strong positive matches.
- Apply a corresponding horizontal-line filter. Because the input contains no horizontal run, that feature map has no comparable positive response.
- Pass the largest responses from both feature maps to a final classifier. The strong vertical response and weak horizontal response support the vertical-line class.
Result: The model classifies the symbol as a vertical line. Reusing each filter at every position lets it detect the same local pattern wherever it appears, while the later classifier combines those detections into the task output.
In a United Nations workplace
At a regional bureau, a records officer is completing the technical evaluation matrix for three bids to classify scanned records before close of business. Each proposed system is tested on the same controlled set of non-sensitive sample pages.
A proposed CNN may reuse learned filters across each page to detect local features such as boxes, stamps and blocks of handwriting, then combine them into larger layout patterns used to classify the document. The evaluation should test those claims consistently, including on rotated, blurred or unfamiliar pages.
Before you try thisBid documents, annexes and procurement-sensitive records must not be uploaded to an external or unapproved tool, as this could compromise confidentiality and the integrity of the evaluation. Use only an authorised environment and cleared test material. Human evaluators must score the bids against the stated criteria rather than accept the model's classifications at face value.
Where you may hear it
- During a procurement unit's technical evaluation of computer-vision bids
- At an ICT governance board review of an image-processing proposal
- In a vendor demonstration of document-image classification
- During a technical briefing on satellite-image analysis
- At a training session on feature maps and image classification
Avoid confusion
Avoid confusion
Often confused with
Recurrent neural network
A recurrent neural network processes a sequence by carrying a changing hidden state from one step to the next, so earlier inputs can influence later outputs.
Transformer
A neural network design that processes many parts of a sequence in parallel by weighting which parts matter most to each other.
Artificial neuron
A small calculation step in a neural network that combines input numbers, adjusts them by learned settings, and passes forward a new number.
Common misconceptions
Common misconception: A CNN looks at the whole image at once in the same way a person might describe a scene.
Correction: Its early layers usually work on small local regions first. Broader context appears only after many local computations are combined across layers. This local-to-global structure is one reason CNNs work well on images, but it also means they can miss information that depends on wider context if the architecture is not designed carefully.
Common misconception: Convolution means the model has built-in edge detectors and does not need much training.
Correction: The mechanism encourages local pattern detection, but the actual filter values are usually learned from training data. Some learned filters may end up behaving like edge detectors, yet which patterns appear depends on the data, objective and architecture.
Common misconception: A CNN is the best neural network for any AI task involving sequences or text.
Correction: CNNs are strongest on grid-like inputs with local spatial structure. They can be used on text or signals in some settings, but recurrent neural networks and especially transformers are often chosen for long-range sequence relationships because they handle context differently.
Quick check
What makes a convolutional neural network different from a basic fully connected network when processing images?
The defining idea is local connectivity plus shared weights: the same filter scans different positions. A basic fully connected layer instead connects inputs broadly, while a recurrent design carries state through a sequence. CNNs do not learn a separate filter for every position.
Go deeper
Go deeper
The simple version
A CNN is a neural network that checks small nearby parts of an image or other grid, finds repeating patterns, and combines those findings into a final result.
The technical version
A CNN is a feed-forward neural network that exploits spatial locality and parameter sharing. In a convolutional layer, a learnable kernel operates over local receptive fields of the input tensor to produce one or more feature maps. The same kernel weights are applied at each spatial position, which reduces parameter count compared with fully connected layers and gives a degree of translation sensitivity or robustness depending on the design. Non-linear activation functions are applied after convolutions, and pooling or strided convolutions may downsample intermediate representations. Stacking layers increases the effective receptive field, allowing later units to respond to more complex spatial patterns assembled from earlier low-level features. For classification, later layers often flatten or aggregate features and map them to class logits, which may then be passed through softmax to produce probability-like outputs. In practice, details such as padding, stride, channel depth, normalisation, and architecture choice affect performance and behaviour.
Reviewed external resources
EncyclopaediaConvolutional neural network (opens in a new tab)Wikipedia
This reference-style overview gives a broad summary of the term, its history, variants and common uses, and suits readers who want a starting point plus terminology.
Intermediatelink checked 2026-07-31
InteractiveTensorFlow Playground (opens in a new tab)TensorFlow
This interactive tool lets readers experiment with simple neural networks in the browser and suits beginners who learn by changing settings, even though it is not a dedicated CNN simulator.
Starterlink checked 2026-07-31
VideoNeural networks, explained visually (opens in a new tab)3Blue1Brown
This visual video builds intuition for neural networks and suits readers who learn best from animated explanations before going deeper into CNN-specific details.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Convolutional neural network (opens in a new tab)YouTube search
This search results page helps readers find many video explanations of convolutional neural networks and suits people who want to compare teaching styles rather than rely on one source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Convolutional neural network (opens in a new tab)Reddit search
This search results page helps readers see community questions and informal explanations about convolutional neural networks and suits readers who want practical discussion with the usual caution about reliability.
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 CNN reuses learned filters across local parts of a grid, then combines the resulting feature maps into larger patterns and a task output.