Beyond Isotropy in JEPAs: Hamiltonian Geometry and Symplectic Prediction
Researchers have introduced HamJEPA, a novel approach to Joint Embedding Predictive Architectures (JEPAs) that moves beyond isotropic regularization. This new method encodes views as phase-space states and uses a learned Hamiltonian leapfrog map for cross-view prediction. Experiments on CIFAR-100 and ImageNet-100 show significant improvements in kNN and linear probe accuracy compared to existing methods like SIGReg. AI
IMPACT Introduces a new method for representation learning that improves performance on downstream tasks.