JEPA Models Revealed as Hidden Markov Models in New arXiv Paper
A recent study posted on arXiv examines the relationship between Joint-Embedding Predictive Architectures (JEPA) and hidden Markov models (HMMs). The researchers unveil that the Predictive Information Bottleneck VJEPA (PIB-VJEPA) shares a computational structure with HMMs, which includes hidden-state inference and Markov chain transitions. They identify four levels of correspondence and establish specific criteria for the HMM equivalence within sequential contexts. Furthermore, the authors propose a novel model called Markov-Chain JEPA (MCJEPA), which incorporates a transition matrix for different applications. This research could significantly impact future developments in the field of representation learning.
Key facts
- Paper on arXiv with ID 2608.13621
- Title: 'Your Probabilistic JEPA Is Secretly a Hidden Markov Model: A State-Space Interpretation of Joint-Embedding Predictive Learning'
- Shows PIB-VJEPA has same computational structure as HMM
- Stochastic context encoder as amortized filtering distribution
- Probabilistic predictor defines latent-state dynamics
- Decoder, inverse target encoder, or induced implicit conditional supplies emission
- Four levels of correspondence distinguished
- Introduces MCJEPA with learned transition matrix
- Matrix powers ensure Chapman–Kolmogorov consistency in finite time-homogeneous case
Entities
Institutions
- arXiv