According to the paper's abstract, LeJEPA—a joint-embedding predictive architecture—derives an isotropic Gaussian embedding target by minimizing a Fisher information criterion under its task assumptions. This criterion is designed to control the task-averaged squared prediction bias, providing a principled foundation for unimodal self-supervised learning.
The new work, MM-LeJEPA, is positioned as a multimodal extension, with the title pointing to "correlated Gaussian regularization" for multimodal joint-embedding predictive architectures. However, the abstract itself focuses on the unimodal LeJEPA derivation and does not provide details about how the correlated Gaussian regularization is implemented or evaluated.
Because only the abstract is available, the article cannot assess the empirical results or the specific multimodal contributions. The theoretical link between Fisher information and Gaussian targets is the clearest takeaway from the source.