Bootstrap Self-Distillation is a self-supervised learning framework that iteratively refines model predictions by treating the model’s own outputs as target supervisory signals, thereby distilling knowledge from an exponentially moving averaged version of the model to stabilize learning and avoid representational collapse.