PEIRA: Learning Predictive Encoders through Inter-View Regressor Alignment

In three linesPEIRA is a non-contrastive self-supervised learning method analyzing JEPA dynamics through a regularized linear regressor. It minimizes an explicit objective based on the trace of the optimal regressor, ensuring stable non-collapsed equilibria aligned with canonical correlation subspaces. Competitive results on ImageNet-1K and CIFAR-10.

Summary generated by Claude — human-verified