Semi-Supervised Learning is a machine-learning paradigm that uses both labelled and unlabelled examples during training, typically when a relatively small amount of labelled data is available alongside a much larger unlabelled dataset. The motivation is that obtaining labels may require expert knowledge, manual annotation, laboratory testing or expensive verification, while collecting raw observations can be comparatively easy. Semi-supervised methods attempt to exploit useful structure in the unlabelled portion while preserving information supplied by labelled examples.
Approaches include consistency regularisation, pseudo-labelling, graph-based methods, co-training and generative techniques. The basic premise is not simply that unlabelled data are always helpful; their usefulness depends on assumptions about how the data are distributed and how their structure relates to the target labels. Poorly chosen or distributionally mismatched unlabelled data can fail to improve performance or can even reinforce errors.
In a conventional supervised setting, the model receives input-target pairs for training examples, whereas a semi-supervised dataset contains only inputs for some observations. A common objective combines a supervised loss calculated on labelled examples with an unsupervised or consistency-based objective calculated using labelled and unlabelled inputs. The relative weighting of these objectives is a hyperparameter or may be changed during training.
Pseudo-labelling is one simple approach: the model generates predicted labels for unlabelled examples and subsequently uses sufficiently confident predictions as additional training targets. This can expand the effective training set, but incorrect pseudo-labels can become self-reinforcing. Confidence thresholds, augmentation-based consistency checks and teacher-student architectures are among techniques used to reduce this risk. Semi-supervised learning therefore introduces an additional feedback loop in which the model's own predictions can influence later learning.
Consistency regularisation is based on the expectation that predictions should remain stable under transformations that preserve the underlying class or task semantics. For example, a model might be encouraged to produce similar predictions for an unlabelled image and an appropriately augmented version of that image. A related approach uses a teacher model to produce targets while a student model learns from them, with the teacher potentially being updated from the student's parameters over time.
Graph-based methods represent examples and their relationships as graphs and use assumptions about nearby or connected observations sharing labels. Co-training uses multiple views or feature sets under conditions where each view can provide complementary information. These approaches embody different assumptions about the geometry and structure of the data, so there is no universal semi-supervised method that is optimal across all datasets.
Semi-supervised learning is particularly useful when labelled examples are expensive but unlabelled examples are abundant. In image analysis, millions of images may be available while only a subset can be manually annotated. In speech processing, large quantities of recorded audio may exist while accurate transcriptions require substantial effort. In industrial applications, sensor measurements can be collected continuously while confirmed failure labels occur relatively rarely. Scientific datasets can contain many observations but only a limited number of experimentally verified classifications. Semi-supervised methods attempt to use the information contained in the unlabelled observations without pretending that those observations have known ground-truth targets. The success of such an approach depends on whether unlabelled examples provide information relevant to the target decision boundary or underlying data structure.
A key theoretical issue is the assumptions required for unlabelled data to improve classification. One commonly discussed assumption is that nearby observations in an appropriate feature space tend to share labels. Another is the cluster assumption, under which decision boundaries should preferentially lie in low-density regions between clusters. A related assumption concerns the relationship between data-generating structure and class membership.
These assumptions can fail when classes overlap substantially, when the most useful discriminating feature is not reflected in overall data geometry, or when unlabelled data come from a different distribution. If a model confidently assigns an incorrect pseudo-label to a region of data, subsequent training can strengthen that error. Semi-supervised learning is therefore not equivalent to receiving free additional labelled data; it uses assumptions about unlabelled observations that must be appropriate for the task.
Semi-supervised learning differs from self-supervised learning because its target task normally retains a conventional labelled component. Self-supervised learning generates supervisory signals directly from the input data, for example by predicting withheld information, whereas semi-supervised learning explicitly has some externally labelled examples and attempts to benefit from additional unlabelled examples. It differs from supervised learning because not all training inputs have external targets. It differs from unsupervised learning because labelled examples provide explicit task information.
It can also incorporate deep learning and transfer learning: a deep neural network may be trained with a semi-supervised objective, and a pretrained representation may be adapted using a mixture of labelled and unlabelled target-domain data. These relationships demonstrate that semi-supervised learning is principally a data-and-objective paradigm rather than a specific architecture.
Advantages include reduced dependence on manual annotation, better utilisation of large unlabelled datasets and the potential to improve task performance when labelled data are scarce. Limitations include sensitivity to incorrect pseudo-labels, distribution mismatch, class imbalance and the validity of structural assumptions about the data. Semi-supervised methods can also increase training complexity because they require multiple objectives, augmentation strategies, teacher models or other machinery.
Evaluation must be performed against a genuinely labelled test set because apparent improvements in an unsupervised consistency objective do not necessarily correspond to better task accuracy. The size and quality of the labelled subset remain important, and a very small or biased labelled sample can cause the model to learn an incorrect interpretation of the unlabelled population. Practical deployment therefore requires monitoring for changes in both labelled and unlabelled data distributions.
Semi-Supervised Learning occupies an intermediate position between conventional supervised and unlabelled-data learning strategies. Its importance has increased as large datasets have become easier to collect while high-quality annotation has remained expensive. In modern neural-network systems, semi-supervised learning may be combined with augmentation, self-training, pretrained representations, self-supervised pretraining and teacher-student methods.
The same dataset can consequently participate in several stages: large unlabelled data may first support self-supervised pretraining, a smaller labelled set may then support supervised adaptation, and unlabelled target-domain data may subsequently contribute through semi-supervised consistency objectives. Federated learning can also be combined with semi-supervised objectives when different participants possess different quantities of labelled and unlabelled local data. The defining feature remains the simultaneous use of labelled and unlabelled examples for the learning process associated with the target task.
