Parallel Variational Monte Carlo Method for Deep State Space Models
A new training method, parallel variational Monte Carlo (PVMC), bridges the gap between auto-encoding and sequential Monte Carlo approaches for deep state space models (DSSMs). PVMC enables robust training for both discriminative and generative tasks, overcoming the poor scalability of classical SMC on modern hardware. The method achieves state-of-the-art results.
Key facts
- arXiv:2605.21108v1
- Announce Type: cross
- PVMC bridges auto-encoding and SMC paradigms
- DSSMs trained via variational lower bound or backpropagating SMC outputs
- Classical SMC forward pass scales poorly on modern hardware
- PVMC enables training for both discriminative and generative tasks
- PVMC achieves state-of-the-art results
Entities
Institutions
- arXiv