New JEPA variant A-JEPA learns causal states from actions

Researchers have developed a new information-theoretic objective for joint-embedding predictive architectures (JEPAs) that aims to recover underlying causal states from observations. This objective combines maximizing conditional likelihood for learning transition dynamics with entropy maximization for preserving latent state information. The work establishes conditions for identifiability, particularly highlighting the importance of sufficient action-induced variation in transition mechanisms. Experiments on synthetic and visual benchmarks demonstrate the effectiveness of this approach, termed action-modulated JEPA (A-JEPA), in recovering states and transferring knowledge to unseen transition dynamics. AI

IMPACT This research could lead to AI models that better understand and predict causal relationships, improving their ability to reason and generalize.

RANK_REASON This is a research paper detailing a new method for learning causal mechanisms within AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• A-JEPA
• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• Influence Flower
• Litmaps
• ScienceCast
• scite Smart Citations

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →