JEPA

What is JEPA?

Joint-embedding predictive architecture that learns by predicting abstract representations rather than reconstructing raw inputs

In other languages

한국어JEPA

원시 입력값 대신 데이터의 표현 공간을 예측하도록 학습하는 자기지도 예측 아키텍처 계열

日本語JEPA

生データそのものではなく、データから得た表現空間で将来や欠けた情報を予測する自己教師あり学習系の手法

Related Terms

• RAGAn architecture that improves LLM accuracy by fetching external knowledge.
• Joint ModelingA technique that processes different types of data such as video and audio simultaneously within a single model to improve coherence.
• Unified ArchitectureA system design where multiple modalities, such as audio and video, are processed by a single, integrated neural network rather than separate, specialized models
• Dense ModelA model architecture where every parameter is utilized for every input token during the forward pass
• Unified Multimodal ModelAI model architecture designed to process and generate multiple data types, such as text and images, within one system