Self-supervised learning representations can suffer from dimensional collapse, but LDReg improves representation quality by regularizing local intrinsic dimensionality.
Regularizing local intrinsic dimensionality improves SSL representation quality.