Self-Supervised Visual Representation Learning

The field studies how visual models can learn useful features without relying on large collections of human-labeled images or videos, enabling transfer to tasks such as classification, segmentation, recognition, and robotic control. Common approaches include contrastive learning across augmented views, masked autoencoders that reconstruct hidden content, self-distillation, and joint-embedding predictive architectures that predict latent features; video and multi-view methods also use temporal and geometric consistency. These methods are evaluated on benchmarks such as ImageNet and COCO, as well as video, 3D, and specialized medical or remote-sensing tasks, with large pretrained encoders often adapted using lightweight probes or fine-tuning. A central tension is how to learn features that are both broadly transferable and sensitive to spatial detail, motion, geometry, or domain-specific cues, while avoiding shortcuts and representation collapse.