SwAV
lastmod 2026-11-10created 2020-07-272 edits
SwAV (Swapping Assignments between Views) is a self-supervised visual representation learning method published by Facebook AI Research. It assigns different augmented views of an image to prototype clusters online and trains to predict one view's cluster assignment from another view, learning features without labels. Unlike contrastive learning, it does not need negative pair comparisons, so it requires neither a memory bank nor large batches, and it also proposed multi-crop augmentation, which mixes in multiple small crops.
Unsupervised Learning of Visual Features by Contrasting Cluster Assignments