If you run self-supervised learning on datasets where each individual image or volume is large, you have to keep your batches fairly small. Konstantinos Almpanakis from our lab found a way to train even in batches of 1, he will present it tomorrow at #ECCV2026!
