DDSSM replaces Gaussian transitions with diffusion for better time series modeling
First model to jointly train autoencoder and diffusion on sequential data.
Traditional deep state space models (SSMs) often assume Gaussian latent transitions, which limits their ability to fit complex, multimodal dynamics. Diffusion models are highly expressive but lack principled inference for underlying system dynamics. The new DDSSM bridges this gap by replacing the Gaussian transition in SSMs with a diffusion model, enabling both accurate forecasting and faithful latent state recovery.
The model introduces a novel joint training objective that simultaneously optimizes the autoencoder and the diffusion transition, extending latent diffusion models to time series. In experiments on a simulated dataset with multimodal transitions, DDSSM outperforms a state-of-the-art deep SSM in both fitting and forecasting. This work, accepted to ProbML 2026, opens new possibilities for modeling complex temporal systems in domains like finance, climate, and neuroscience.
- Replaces Gaussian latent transitions in state space models with a diffusion model for better expressiveness.
- Solves the open problem of jointly training an autoencoder and a diffusion model on sequential data.
- Outperforms a state-of-the-art deep SSM on a simulated multimodal time series in both fit and forecast accuracy.
Why It Matters
Enables more accurate forecasting and deeper understanding of complex, non-Gaussian temporal systems.