semi-supervised learning

Gold definitionUpdated Apr 2, 2026

Definition

Semi-supervised learning (SSL) is a machine learning paradigm that leverages both a small amount of labeled data and a large amount of unlabeled data for training. It aims to improve model performance and generalization, especially in low-data scenarios, by inferring useful information from the abundant unlabeled examples.

At a glance

Executive summary

Semi-supervised learning helps AI models learn effectively even when only a small portion of the training data is labeled. It combines the small labeled dataset with a large amount of unlabeled data, using various techniques to infer useful information from the unlabeled examples to improve accuracy and generalization.

TL;DR

Semi-supervised learning trains AI models using a mix of a little bit of labeled data and a lot of unlabeled data to make them smarter and more accurate.

Key points

Combines labeled and unlabeled data, often by generating pseudo-labels or optimizing for data coherence.
Overcomes the high cost and scarcity of labeled data, especially in low-data scenarios.
Used by researchers in NLP (e.g., LMs), computer vision, drug discovery, and material design.
Unlike supervised learning which needs all data labeled, SSL uses both to achieve better performance with less annotation effort.
Growing interest in theoretical understanding of self-improvement mechanisms and robust pseudo-labeling for diverse tasks.

Use cases

Molecular Property Prediction: Predicting properties of new drug candidates or materials using limited experimental data and vast databases of uncharacterized molecules (e.g., SemiMol).
Image Classification: Training robust image classifiers for rare diseases or specialized domains where only a few annotated images are available, supplemented by many unannotated ones.
Natural Language Processing: Improving language model performance and generalization by leveraging massive amounts of raw text data alongside smaller, task-specific labeled datasets.
Speech Recognition: Enhancing speech recognition systems by training on limited transcribed audio and abundant untranscribed audio.
Fraud Detection: Identifying fraudulent transactions with few confirmed fraud examples by utilizing patterns from a large volume of normal, unlabeled transactions.

Also known as

SSL, self-training, pseudo-labeling, consistency regularization, transductive learning, inductive learning (in SSL context)