Learn AI /How Machines Learn /Self-supervised learning
Self-supervised learning
- Intermediate
- How Machines Learn
- Technical readers
Back to the milestone you came from
In five seconds
What it is
A training approach that creates its own learning targets from raw data, so a model can learn useful patterns without manual labels.
In plain words
Instead of asking people to label every example, this method hides or rearranges part of the data and trains the model to predict the missing or related part.
An everyday example
A team is preparing a long report in English, Arabic and French. During practice, one person covers a few words in each sentence and asks the others to fill them back in using the surrounding context. No one has added topic labels such as “health” or “education”, but the exercise still teaches people the structure, vocabulary and patterns of the text. Later, those people can sort new paragraphs by topic faster because they have already learned how the language in the report tends to fit together.
People in the story already know language and world facts, while a model starts with parameters and adjusts them through numerical optimisation. Also, self-supervised learning is not limited to filling in blanks, and many methods use image crops, audio segments or other transformations rather than hidden words.
Remember thisSelf-supervised learning creates prediction targets from the raw data itself, allowing a model to learn reusable representations before a smaller labelled dataset is applied.
This term is contested
The boundary between self-supervised learning and unsupervised learning is described differently by different authors. Many treat self-supervised learning as a subset of unsupervised learning, while others present it as a distinct category because it uses explicit automatically generated targets.
Understand it
Understand it
Why it matters
For decision-makers, self-supervised learning matters because labelled data is expensive, slow and sometimes impossible to obtain at scale. This approach can use large amounts of existing raw data to learn useful patterns first, then reduce how much manual labelling is needed for a later task. It can improve performance in settings where organisations have many documents, images or signals but limited expert time. It also changes procurement and governance questions: you may need to ask not only about labels, but about the source data, the automatic training task, privacy controls, and whether the learned representations transfer well to the real task.
How self-supervised learning creates its own training signal
Show how raw data can be turned into an automatic prediction task, producing learned representations that can later support another task.
01
Raw data
Start with unlabelled data such as documents, images, audio clips or time series.
Draw a stack of mixed data cards on the left labelled “raw unlabelled data”. Show small icons for text, image and waveform.
02
Create a training task
Transform each example so one part becomes the input and another part becomes the target to predict.
Show a sentence with one word masked, an image with one patch hidden, and a waveform with a short muted segment. Add a bracket labelled “target comes from the same example”.
03
Train the model
The model predicts the missing, matching or next part, compares prediction with the automatic target, and updates its parameters.
Draw a model box with an input arrow in, a prediction arrow out, and a feedback loop labelled “error → parameter update”.
04
Reuse what was learned
The trained model or its representations can then support a later task with fewer human labels or better performance.
Show outputs feeding into small boxes labelled “classification”, “search” and “anomaly detection”, with a note “often fine-tuned later”.
A left-to-right flow with four boxes. Read from raw data, to an automatic transformation that creates an input-target pair, to model training, to learned representations and later reuse. Arrows show that the target comes from the same original data rather than from a human label.
Hovering over each transformation example highlights the matching automatic target and shows that no manual label was added.
Describe this without the diagram
The diagram is a four-stage flow. First, raw unlabelled data such as text, images and audio enter the process. Second, each example is automatically transformed so that one part is hidden, removed, paired or reordered, creating a prediction target from the same original example. Third, a model receives the modified input, predicts the missing or related part, measures error against the automatic target, and updates its parameters. Fourth, the result is a model with learned representations that can later be adapted for tasks like classification or search, often using fewer manual labels.
With reduced motion: A static four-box flow replaces all animation. Arrows connect raw data, automatic target creation, model training with feedback, and later reuse for downstream tasks.
In many machine learning tasks, people first attach labels to examples, such as “spam” or “not spam”. Self-supervised learning uses a different idea: it builds a prediction task from the data itself. The input is still ordinary data such as text, images, audio or sensor readings, but the target to predict is generated automatically rather than written by a human.
A simple example is text. If you hide one word in a sentence and ask a model to predict the missing word, the sentence itself provides the training signal. In images, you might hide part of an image or ask whether two cropped views came from the same original image. These are called pretext tasks: training tasks designed to make the model learn useful internal patterns.
What changes inside the model during training are its parameters, the adjustable numbers that shape its predictions. The model compares its prediction with the automatically created target, measures the error, and updates those parameters many times across many training examples. Over time, it tends to build internal representations: compact patterns that capture structure in the data.
What comes out is usually not a final answer to a business or policy question by itself. Often the main output is a model that has learned representations that can later be adapted for another task, such as classification, search, summarisation or anomaly detection. This is why self-supervised learning is often used before supervised learning or transfer learning.
It is often confused with unsupervised learning, but they are not identical. In self-supervised learning, there is still a target to predict; it is just derived from the input data instead of supplied as a separate human label.
How “Covering words in a report draft” maps onto the real thing
| In the story | In the AI system |
|---|---|
| covered words in sentences | parts of the input hidden or transformed to create a prediction target |
| using surrounding words to fill gaps | learning from context within the same training example |
| practice before the real sorting task | pretraining before adapting to a downstream task |
| getting better at recognising document patterns | learning internal representations that capture structure |
A worked example
Pretraining a language model on public transport updates using masked-word prediction
Starting from: Original sentence: “Bus 42 is delayed by 15 minutes due to heavy rain near the central station.”
- Replace one word with a mask to create the model input: “Bus 42 is delayed by 15 minutes due to heavy rain near the central [MASK].”
- Use the hidden word, “station”, as the target. It comes directly from the original sentence, rather than from a human annotation.
- Repeat this transformation across many updates. Each time, adjust the model when its prediction differs from the hidden word.
- Use the resulting representations for a later task, such as classifying updates as delays, cancellations or route changes, with a smaller set of labelled examples.
Result: The pretraining data had no topic labels, but the masking task provided its own targets. By learning to predict missing words, the model may capture language patterns that are useful for the later classification task.
In a United Nations workplace
A records officer in the Resident Coordinator's Office is drafting the note for the file after a bilateral held by the data working group of the UN Country Team. Before close of business, the officer must record a proposal to pretrain an internal search model on cleared workplans and progress reports, without first tagging every document by topic.
During pretraining, the model could create targets by masking passages in each document and learning to reconstruct them. This may produce representations that support later retrieval or grouping with fewer manually tagged examples.
Before you try thisClearance for circulation does not automatically permit use as training data. Personal, confidential, protection-sensitive or security-sensitive material must not be entered into a system that is not approved for it, because training can expose or reproduce source content. Confirm the authorised purpose, access controls, retention arrangements and human review before proceeding.
Where you may hear it
- A vendor demo during an ICT governance board meeting
- A technical review of an internal search proposal
- An inter-agency working group discussing document retrieval
- A procurement clarification on proposed training data
- A research briefing on foundation-model pretraining
Avoid confusion
Avoid confusion
Often confused with
Unsupervised learning
A way of training a model on unlabelled data to identify possible structure, similarities, groups or unusual cases without target answers for each example.
Semi-supervised learning
A learning approach that combines labelled and unlabelled examples to train a predictor for the same target task.
Supervised learning
A method for training a model on examples paired with target outputs, so it can predict outputs for new inputs.
Common misconceptions
Common misconception: Self-supervised learning means the model learns with no targets at all.
Correction: It usually does have targets, but they are created from the data itself. For example, a hidden word in a sentence can be the target. That is different from having no prediction objective.
Common misconception: It is just another name for unsupervised learning.
Correction: The terms overlap in casual speech, but they are not the same. Unsupervised learning often focuses on finding structure such as clusters without explicit target prediction. Self-supervised learning typically defines a prediction task with automatically generated targets.
Common misconception: If a model is trained this way, it no longer needs labelled data.
Correction: Sometimes self-supervised pretraining is enough for useful representations, but many practical applications still benefit from supervised fine-tuning or evaluation on labelled data. The method often reduces label needs rather than removing them entirely.
Quick check
Which statement best describes self-supervised learning?
Self-supervised learning usually sets up a prediction problem where the target comes from the same raw example, such as a masked word or missing image patch. That differs from supervised learning with human labels, from clustering-only approaches, and from reinforcement learning with rewards.
Go deeper
Go deeper
The simple version
Instead of asking people to label every example, this method hides or rearranges part of the data and trains the model to predict the missing or related part.
The technical version
Self-supervised learning trains on unlabelled data by deriving supervisory signals from each example itself. A transformation defines an input-target pair, such as masked-token prediction in text, contrastive pairing of two augmented views of the same image, next-step prediction in sequences, or reconstruction of corrupted inputs. The optimisation objective updates model parameters so internal representations capture statistical structure that is useful beyond the pretext task. Those representations can then be transferred, fine-tuned, or used directly for downstream tasks. The key distinction from standard supervised learning is that the target is not an external human-provided label, while the distinction from generic unsupervised learning is that training commonly still uses an explicit predictive loss.
Reviewed external resources
EncyclopaediaSelf-supervised learning (opens in a new tab)Wikipedia
This link gives a broad overview of the term, its history and related methods, and suits readers who want a general reference before going deeper.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course link suits readers who want practical machine learning foundations that help place self-supervised learning in the wider training workflow.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This video source suits readers who prefer step-by-step spoken explanations and intuition before reading more technical material.
Starterlink checked 2026-07-31
VideoVideo explanations: Self-supervised learning (opens in a new tab)YouTube search
This search results page helps readers compare several video explanations of self-supervised learning, and suits people choosing a format or level that fits them.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Self-supervised learning (opens in a new tab)Reddit search
This search results page lets readers see how practitioners and learners discuss the term informally, and suits readers who want common questions, examples and debates rather than a single curated explanation.
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
Self-supervised learning creates prediction targets from the raw data itself, allowing a model to learn reusable representations before a smaller labelled dataset is applied.