P-JEPA: Procedural Video Representation Learning via Joint Embedding Predictive Architecture

The article presents the Procedural Joint Embedding Predictive Architecture (P-JEPA), which enables long-duration video representation learning by addressing the limitations of self-attention in processing procedural videos with long-range dependencies. P-JEPA effectively ingests videos over 30 minutes, leveraging a dense, frame-aligned action space and pooled masked latent vector predictions, and demonstrates state-of-the-art performance in action classification on the EgoExo4D dataset while utilizing significantly fewer parameters than LLM-based approaches. This advancement is crucial for practitioners developing intelligent assistance systems that require understanding complex, multi-step tasks in real-time.