Semi-supervised learning is a training paradigm that combines a small labelled dataset with a much larger pool of unlabelled data, sitting between purely supervised learning (all examples labelled) and unsupervised learning (no labels at all). The model first learns an initial representation from the labelled examples, then uses that representation to generate pseudo-labels or consistency constraints on the unlabelled data, effectively expanding the training signal without the cost of manual annotation. Techniques such as pseudo-labelling, label propagation, and consistency regularisation (for example, FixMatch) are common implementations.

Think of it like a student who studies ten worked exam answers in detail, then practises on hundreds of unmarked practice papers by checking whether their own answers are internally consistent and similar to the solved examples — self-correcting as they go.

# Pseudo-labelling sketch
supervised_model.fit(X_labelled, y_labelled)
pseudo_labels = supervised_model.predict_proba(X_unlabelled)
high_conf = pseudo_labels.max(axis=1) > 0.95   # threshold
X_combined = np.vstack([X_labelled, X_unlabelled[high_conf]])
y_combined = np.hstack([y_labelled,
                        pseudo_labels[high_conf].argmax(1)])
final_model.fit(X_combined, y_combined)

Semi-supervised learning bridges Pretraining (which learns from vast unlabelled corpora) and supervised Fine-tuning; it is related to Overfitting concerns because adding noisy pseudo-labels can introduce new error sources. Understanding Loss Function design matters here, as consistency losses on unlabelled data must be weighted carefully against the supervised signal.

You will encounter semi-supervised learning when annotation budgets are tight — common in medical imaging, rare-language NLP, and specialised industrial domains where labelling requires expensive domain expertise.