SplitJEPA: Learning Invariant and Variant Latent Worlds without Reconstruction
SplitJEPA claims decoder-free recovery of both shared and changing latent factors.
The paper introduces a JEPA-style method that separates invariant and variant subspaces directly in representation space. Its proof covers stationary Gaussian predictive dynamics with a full-rank variation condition, identifying the two blocks up to independent isometries. The authors say experiments on synthetic nonlinear systems and robotic manipulation tasks support gains in robustness and efficiency. Source: ArXiv · AI/CL/LG's note.
The paper introduces a JEPA-style method that separates invariant and variant subspaces directly in representation space. Its proof covers stationary Gaussian predictive dynamics with a full-rank variation condition, identifying the two blocks up to independent isometries. The authors say experiments on synthetic nonlinear systems and robotic manipulation tasks support gains in robustness and efficiency. Source: ArXiv · AI/CL/LG's note.
score 4