Скачать или смотреть VL-JEPA: Joint Embedding Predictive Architecture for Vision-Language. Vision Language Models (VLMs)

VL-JEPA: Joint Embedding Predictive Architecture for Vision-Language

The podcast provides the technical overview of the VL-JEPA (Vision-Language Joint Embedding Predictive Architecture), a novel, non-generative model designed for vision-language tasks, offering a paradigm shift from traditional autoregressive Vision Language Models (VLMs). VL-JEPA predicts continuous embeddings of target texts in an abstract semantic space, enabling it to achieve higher performance and training efficiency with fewer trainable parameters compared to token-space VLMs in controlled comparisons. This architecture is particularly advantageous for real-time and streaming video applications due to its non-autoregressive nature and native support for selective decoding, which significantly reduces inference time and latency. The resulting model is highly versatile, demonstrating strong performance across a unified architecture on tasks including VQA, classification, and text-to-video retrieval.