DES-JEPA Learns Shared Vocabularies for High-Dimensional Event Sequences

DES-JEPA uses a joint-embedding predictive architecture to learn a shared discrete vocabulary from stochastic processes. This addresses the limitation of existing foundation models that are restricted to small, fixed sets of event types in high-dimensional discrete event sequences like clinical logs.