JEPA models shown to be equivalent to Hidden Markov Models

A new research paper proposes that Joint-Embedding Predictive Learning (JEPA) models, when fully time-indexed, exhibit the same computational structure as Hidden Markov Models (HMMs). The paper details how components of JEPA, such as the stochastic context encoder and probabilistic predictor, correspond to the inference, propagation, and emission roles found in HMMs. To solidify this connection, the researchers introduce Markov-Chain JEPA (MCJEPA), which uses a learned transition matrix to ensure consistency with Chapman-Kolmogorov equations. Experiments support the interpretation of JEPA's predictive learning as seeking a compact predictive state, distinguishing it from traditional HMM sequence learning. AI

IMPACT Provides a new theoretical framework for understanding and potentially improving predictive learning models in AI.

RANK_REASON Academic paper published on arXiv detailing a theoretical connection between two AI modeling approaches. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Chapman--Kolmogorov
• hidden Markov model
• Hugging Face
• Joint-Embedding Predictive Learning
• Markov-Chain JEPA
• Predictive Information Bottleneck VJEPA
• Yongchao Huang

• paper
• model release

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