noun
Self-supervised learning is a machine learning technique where models train on unlabeled data by generating their own labels from the input. Unlike supervised learning, which requires human-annotated data, self-supervised models rely on inherent structures in the data to create training signals.
For example, a model might learn to predict missing words in a sentence or reconstruct scrambled images. This approach is particularly useful for tasks like natural language processing and image recognition, where labeled data is scarce or expensive to obtain.
For developers, self-supervised learning reduces the need for manual labeling, lowering costs and speeding up model training. However, it requires large datasets and sophisticated algorithms to ensure the generated labels are meaningful. A real-world example is Meta’s DALL·E 2, which uses self-supervised learning to generate high-quality images from text prompts, achieving results comparable to human-created art.
A self-supervised model might learn to colorize black-and-white photos by predicting missing color values from surrounding pixels.
In other languages
See also: Supervised learning · Unsupervised learning · Transfer Learning