Joint-embedding systems can collapse by ignoring inputs and producing constant representations; preventing collapse is the central problem in self-supervised learning for these systems and requires either contrastive methods (pushing up energy of non-data points) or regularized methods (minimizing volume of low-energy regions)