Supervised learning is a category of Machine learning in which a model is trained on a dataset of examples that each pair an input with a known correct output, or label, and the model's objective is to learn a general mapping from inputs to outputs that generalizes to new, unseen examples. The word "supervised" refers to the fact that a labeled answer key, typically produced by human annotators or gathered from historical records, supervises the learning process by telling the model exactly what the correct answer should have been for each training example.
Core tasks
Supervised learning problems fall broadly into two categories. Classification tasks predict a discrete category, such as whether an email is spam, what object appears in an image, or what sentiment a review expresses. Regression tasks predict a continuous numerical value, such as a house price or a temperature forecast. Both are trained the same way: the model makes a prediction, a Loss function measures how far that prediction is from the true label, and Gradient descent, computed via Backpropagation in the case of a Neural network, adjusts the model's parameters to reduce that error, repeated over many examples and passes through the data.
History
Supervised learning predates deep learning by decades and includes classical algorithms such as linear and logistic regression, decision trees, support vector machines, developed and formalized in large part by Vladimir Vapnik, and k-nearest neighbors. The technique gained enormous momentum with the availability of large labeled datasets and cheap compute; the 2012 ImageNet competition, where AlexNet dramatically outperformed prior methods on a large labeled image dataset, is widely credited with demonstrating that supervised deep learning could scale to outperform hand-engineered approaches, triggering the broader Deep learning resurgence.
Data and labeling
Supervised learning's central practical bottleneck is the need for labeled data, which is often expensive and slow to produce, requiring human annotators to categorize, transcribe, or otherwise mark up large numbers of examples. This constraint motivated the development of alternative paradigms that reduce or eliminate the need for labels, including Unsupervised learning, which finds structure in unlabeled data, and Self-supervised learning, which generates its own training signal from unlabeled data, an approach that underlies the Pretraining of modern large language models. Supervised learning remains essential at other stages of the modern AI pipeline, however, most notably in supervised Fine-tuning, where a pretrained model is adapted using labeled examples of desired input-output behavior, a step used in turning a base language model into an instruction-following assistant.
Evaluation
A model's performance in supervised learning is typically measured on a held-out test set that was not used during training, in order to estimate how well it generalizes rather than merely memorizing its training data, a failure mode called Overfitting. Common evaluation metrics include accuracy, precision, recall, and F1 score for classification, and mean squared error for regression, with the appropriate metric depending heavily on the task and the relative cost of different kinds of mistakes, such as false positives versus false negatives in a medical diagnosis setting.
Significance and limitations
Supervised learning has been the dominant paradigm behind most deployed machine learning systems for decades, from spam filters and credit scoring to medical image diagnosis and speech recognition. Its central limitation is a strict dependence on the quality and coverage of its labeled training data: a model can only be as good as its labels, and labels that reflect historical human bias or incomplete coverage of real-world scenarios will produce a model that inherits or amplifies those flaws, a concern central to research on Algorithmic bias. As unlabeled and self-supervised approaches have matured, supervised learning has increasingly become one stage in a larger pipeline, rather than the entire training process for the most capable modern AI systems.