Semi-supervised learning is a machine-learning approach that uses a combination of labelled and unlabelled data during model development. The labelled subset provides explicit target information, while the much larger unlabelled subset can contribute information about the structure or distribution of the input data. It is particularly useful when obtaining labels requires substantial human effort, specialised expertise, expensive experiments or access to information that is difficult to collect.
A semi-supervised algorithm may combine a conventional supervised loss on labelled examples with an additional objective derived from unlabelled examples. The exact mechanism varies widely and can include pseudo-labelling, consistency regularisation, graph-based propagation, generative modelling, co-training or other techniques. Semi-supervised learning therefore occupies a position between conventional supervised learning, which relies on labelled examples for its primary training signal, and methods that learn without externally supplied labels.
The basic mathematical setting contains a labelled dataset L = {(x_i, y_i)} and an unlabelled dataset U = {x_j}. A model is trained using the known targets in L, while information from U is incorporated through assumptions about how examples should behave. One common strategy is to minimise a combined objective such as a supervised loss plus an unsupervised or consistency loss.
The supervised component encourages correct predictions on labelled examples; the additional component may encourage predictions to remain stable under perturbations, encourage unlabelled examples to lie in plausible decision regions, or use model-generated pseudo-labels as provisional targets. The relative weighting of these objectives is a crucial hyperparameter because an overly strong unlabelled-data objective can amplify incorrect assumptions. Semi-supervised learning is therefore not simply “training on some labelled data and some unlabelled data” in a generic sense; the unlabelled examples must participate through a defined learning mechanism.
Pseudo-labelling is a widely used semi-supervised technique. A model first makes predictions on unlabelled examples, and predictions with sufficiently high confidence can be treated as provisional labels for additional training. The model can then learn from these pseudo-labelled examples alongside the original labelled data. This can effectively expand the training set, but it introduces a risk of confirmation bias: incorrect predictions can become training targets and reinforce the model's own mistakes.
Modern methods attempt to reduce this problem using confidence thresholds, teacher-student architectures, calibrated uncertainty, strong and weak data augmentations, consistency constraints and other safeguards. The quality of pseudo-labels depends on the model's initial performance, the distribution of unlabelled data and the reliability of its confidence estimates. Pseudo-labelling therefore works best when the model has sufficient initial competence and the unlabelled data is relevant to the target task.
Consistency regularisation is another important family of methods. The central idea is that a model's prediction should remain reasonably stable when the same unlabelled example is subjected to transformations that should not change its semantic identity. For an image, this might involve cropping, flipping or other suitable augmentation; for text, appropriate perturbations require greater care because small changes can alter meaning. A model can be penalised when predictions differ excessively between different views of the same unlabelled example.
Teacher-student methods extend this idea by having one model produce targets or guidance for another, sometimes using an averaged or slowly changing teacher. These approaches exploit the assumption that meaningful decision boundaries should respect the underlying structure of the data. The assumption is useful but not universally valid, and inappropriate augmentations or distributional assumptions can harm performance.
Semi-supervised learning has particular value in domains where unlabelled data is plentiful but labels are scarce. Medical imaging can produce many scans while expert annotation requires specialist clinicians; speech applications can have large audio collections but relatively limited transcriptions; image and video repositories can contain enormous numbers of examples without object-level annotations; and industrial sensors can produce continuous measurements while failure labels remain rare.
Semi-supervised methods can use the inexpensive unlabelled portion to improve representation and decision boundaries while relying on a smaller labelled set for task-specific grounding. However, the unlabelled data must generally be relevant to the deployment distribution. If it contains very different populations, systematic artefacts or misleading patterns, using it can reduce rather than improve performance.
Semi-supervised learning should be distinguished from self-supervised learning. In self-supervised learning, the training target is generated from the input data itself, such as predicting masked words or withheld image information, and the resulting representation may later be fine-tuned using labelled data. In semi-supervised learning, there is explicitly a mixture of examples with externally supplied labels and examples without them in the learning procedure.
The distinction can become blurred in modern pipelines because a system may use self-supervised pretraining followed by semi-supervised fine-tuning. Transfer learning is related but describes reusing knowledge from one task or dataset for another rather than specifically describing the proportion of labelled data. Active learning is also related: it attempts to select which unlabelled examples should be labelled by a human or other oracle, thereby concentrating annotation effort where it is expected to provide the greatest value.
The effectiveness of semi-supervised learning depends on assumptions about the relationship between labelled and unlabelled data. A common assumption is that examples close to each other in an appropriate representation should tend to have similar labels, while another is that the decision boundary should avoid regions of high data density. These assumptions can fail when different classes overlap substantially, when rare classes are important, or when the unlabelled dataset contains categories absent from the labelled data.
Dataset imbalance can be particularly problematic because pseudo-labelling may preferentially reinforce dominant classes. Evaluation must therefore compare against supervised baselines using the same labelled data and should test performance on a held-out set representative of actual deployment. Merely improving agreement with pseudo-labels is not evidence of improved real-world accuracy because the pseudo-labels may themselves be wrong.
Semi-supervised learning is consequently a strategy for extracting additional value from unlabelled examples while retaining the grounding supplied by labelled examples. Its main benefit is reduced dependence on expensive annotation, while its main risk is that assumptions about the unlabelled data can introduce or amplify errors. Successful systems often combine carefully selected labelled data, representative unlabelled data, robust augmentation, confidence control, consistency objectives and rigorous evaluation.
It is particularly useful when the cost of acquiring labels dominates the cost of collecting raw observations. In modern AI, semi-supervised methods also form part of broader training pipelines that combine self-supervised pretraining, supervised fine-tuning, synthetic data, human feedback and other learning signals. The term therefore identifies the use of both labelled and unlabelled data in learning, while the exact algorithm determines how each source contributes to the resulting model.
