Unsupervised Learning

Using only unlabeled data and letting the model discover structure in it on its own.

Unsupervised learning is the machine-learning paradigm contrasted with supervised learning: the training data has only inputs, no human-annotated answers, and the model must find patterns on its own. Classic tasks include clustering (such as k-means, which groups similar samples together), dimensionality reduction (such as principal component analysis, PCA), and density estimation (learning the data's underlying probability distribution); generative models like autoencoders are also often grouped here. Self-supervised learning (constructing a supervisory signal from the data itself, such as predicting a hidden part) is sometimes considered a branch of it. It matters because labeling is expensive while unlabeled data is abundant, and large-model pretraining relies heavily on unlabeled text, images, and video. In embodied AI, learning latent actions from action-free human video, and letting a robot discover a variety of behaviors through unsupervised skill discovery, both follow this line of thinking.

ExampleRunning k-means clustering on a batch of unlabeled tabletop scene images: the algorithm doesn't know any category names, it just groups the images by feature similarity.

Related

Supervised Learning · Self-Supervised Learning · Representation Learning · Pre-training · Unsupervised Skill Discovery · Autoencoder

Sources

Wikipedia: Unsupervised learning