SplitJEPA: Learning Invariant and Variant Latent Worlds without Reconstruction
SplitJEPA introduces a reconstruction-free Joint Embedding Predictive Architecture (JEPA) to jointly recover both the latent state and its invariant and variant organization directly in representation space. It proves that under specific dynamics, this method identifies these invariant and variant subspaces up to block-wise isometries without needing an observation decoder.