presentsUNAIVERSE
The UN AI timeline

    Learn AI /AI Foundations /Deep learning

    Deep learning

    Detail

    In five seconds

    What it is

    A method that trains many-layered computer models to find useful patterns in large amounts of data.

    In plain words

    It is a way of training a computer system with many layers so it can learn useful patterns from lots of examples.

    An everyday example

    In a regional office, staff receive thousands of photographed forms from phones and scanners. A new clerk first notices simple things like where the logo sits, whether there is a barcode, and how many boxes appear near the top. A more experienced clerk then combines those clues to tell whether the image is a registration form, an expense claim or an attendance sheet. After checking many labelled examples, the team gets faster and makes fewer mistakes on new form photos.

    Where the story stops working

    People can explain their reasoning in words and can use broad office knowledge after only a few examples. Deep learning models usually need many examples, learn through numerical adjustment rather than human-style reasoning, and may be hard to explain clearly.

    Remember thisDeep learning trains many-layer neural networks by repeatedly adjusting weights to reduce error. The layers can learn increasingly complex features, but a plausible output is not proof that the model understood the material.

    This term is contested

    The exact boundary between deep learning, neural networks and broader machine learning usage is not always drawn the same way across textbooks, companies and researchers.

    Understand it

    Understand it

    Why it matters

    Non-specialists hear this term often because many modern AI systems depend on it. Knowing the basics helps you ask practical questions: what data trained it, what task it suits, how much oversight it needs, and where it may fail.

    How deep learning turns input into output

    After viewing this, the reader should be able to say that deep learning passes data through many layers, compares the result with a target during training, and adjusts weights to improve later outputs.

    1. 01

      1. Input data

      The process starts with data such as an image, sound clip, sentence or table of numbers.

      Draw a box labelled 'Input' with small icons for image, audio waveform, text lines and numbers.

    2. 02

      2. Early layers

      The first layers pick up simple patterns, such as edges, shapes, sound fragments or short word relationships.

      Draw the first layer as a column of circles with tiny edge and line symbols above it.

    3. 03

      3. Deeper layers

      Later layers combine simpler patterns into more useful structures for the task.

      Draw two more layer columns; annotate with examples like 'form layout', 'spoken word pattern' or 'phrase structure'.

    4. 04

      4. Output

      The model produces a result such as a label, score, translation or generated content.

      Draw an output box with example labels like 'template A', 'French translation', or '0.87 risk score'.

    5. 05

      5. Training feedback

      During training, the output is compared with the known answer, the error is measured, and the weights are adjusted.

      Add a target box near the output, an 'error' marker between output and target, and a curved arrow back to the layers labelled 'adjust weights'.

    Show a left-to-right flow with one input, three hidden layers, one output, and a training feedback arrow from error back to the layers. Each layer should show patterns becoming more complex.

    Hovering or tapping a layer highlights its role and updates a caption from simple patterns to more complex patterns; a toggle switches between image, speech and text examples.

    Describe this without the diagram

    A flow diagram reads from left to right. First is input data: an image, audio, text or numbers. Next come early layers, which detect simple patterns. Then deeper layers combine those simple patterns into more complex structures related to the task. The output stage produces a label, score, translation or generated content. A final training feedback path compares the output with a known target, measures error, and sends an instruction back through the layers to adjust weights so later outputs improve.

    With reduced motion: A static left-to-right diagram shows input, early layers, deeper layers, output, and a separate curved arrow labelled 'training feedback adjusts weights after comparing with target'.

    Deep learning is a type of machine learning that uses neural networks with many layers. These mathematical systems turn inputs into outputs. The input might be an image, sound, text or numbers. The output might be a label, score, translation or newly generated content.

    What makes it “deep” is that the neural network has many layers. Each layer transforms values from the previous layer and passes the results on. Early layers may pick up simple patterns, such as edges in an image or short word patterns in text, while later layers combine those into more complex patterns.

    During training, an algorithm adjusts many internal values, often called weights, to improve performance on a training objective. In supervised training, for example, it compares the model’s output with an expected answer, measures the error and updates the weights.

    A deep learning model might classify medical images, recognise speech, translate between languages or generate text and images. Deep learning is a type of machine learning. It uses neural networks, but not every neural network is deep.

    How “Sorting scanned forms at a busy office” maps onto the real thing

    In the storyIn the AI system
    simple visual clues like logo position and boxesearly layers detecting basic patterns
    experienced clerk combining clueslater layers combining simple features into more complex ones
    many labelled form examplestraining data in a dataset
    team improving after checking past mistakesadjusting weights to reduce error during training

    A worked example

    A reporting officer at a regional bureau is preparing a briefing note for morning clearance. Image attachments from incoming situation reports must be separated into maps, tables and narrative pages so the relevant material reaches the correct section.

    Starting from: A large, approved collection of document images labelled as maps, tables or narrative pages, with a separate set reserved for testing.

    1. The images are standardised and divided into training and test sets.
    2. A multi-layer neural network receives each training image and produces a predicted document category.
    3. The prediction is compared with the known label. The model then adjusts its trainable weights to reduce the error and repeats this process across the training data.
    4. Across its layers, the model learns progressively more complex features, from edges and blocks of text to page layouts associated with maps, tables or narrative pages.
    5. The reporting officer tests the trained model on images it did not see during training and manually reviews uncertain or unusual pages.

    Result: The model routes familiar page types more quickly, while the reporting officer retains responsibility for misclassified, low-confidence and novel layouts before the briefing note goes for clearance.

    In a United Nations workplace

    A programme officer in a country office is clearing an implementing partner’s progress report by close of business. Its annexes arrive as mixed scans of delivery records, invoices and attendance sheets. An approved deep learning tool sorts the pages into the correct review queues before the officer checks the package and uploads it to the portal.

    A many-layer neural network can learn visual features from a large, labelled archive of document images. During training, its weights are repeatedly adjusted to reduce classification errors. Earlier layers may detect lines and shapes, while later layers combine them into layouts associated with particular document types.

    Before you try thisInvoices and attendance sheets can contain personal and procurement-sensitive information. They must not be uploaded to any tool not approved for those data, because disclosure could harm individuals or compromise a procurement process. Even in an authorised environment, minimise the data and require staff to confirm the routing. The model may recognise a familiar layout without establishing that the evidence is genuine or sufficient.

    Where you may hear it

    • A vendor demonstration during a procurement evaluation
    • An ICT governance board reviewing an image-analysis pilot
    • An inter-agency working group discussing speech recognition
    • A Headquarters briefing on automated translation

    Avoid confusion

    Avoid confusion

    Often confused with

    Side-by-side

    Common misconceptions

    • Deep learning and machine learning are different unrelated fields.

      Deep learning is a subset of machine learning. It is one family of methods within the broader area, usually used when layered models learn patterns from data.

    • More layers always mean better results.

      Extra depth can help with complex tasks, but performance also depends on data quality, model design, computing resources and careful training. A deeper model can also be harder to train or easier to misuse.

    • A deep learning model understands content like a person does.

      It finds statistical patterns that help produce useful outputs. That can look impressive, but the process is not the same as human understanding, and the model can still fail in unexpected ways.

    • Deep learning only works for images.

      It is widely used for images, but also for speech, text, translation, forecasting and generation. The key idea is layered pattern learning, not one specific media type.

    Quick check

    Which statement best describes deep learning?

    Go deeper

    Go deeper

    The simple version

    It is a way of training a computer system with many layers so it can learn useful patterns from lots of examples.

    The technical version

    It typically uses multi-layer neural networks with trainable weights. Training iteratively updates those weights to minimise error on example data, allowing hierarchical feature learning from raw or lightly processed inputs.

    Reviewed external resources

    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

    Deep learning trains many-layer neural networks by repeatedly adjusting weights to reduce error. The layers can learn increasingly complex features, but a plausible output is not proof that the model understood the material.