Scan-Parallel Variational Inference for event-driven Latent Jump Models
Continuous-time latent state-space models are widely used for irregularly observed time series, typically representing uncertainty through a random initial state followed by smooth ordinary differential equation (ODE) dynamics, or Brownian perturbations accumulated by a stochastic differential…
This paper introduces latent jump models (LJMs) and a novel training method called Scan-Parallel Variational Inference (SPVI) for improving the efficiency of continuous-time latent state-space models in handling irregular time series data.
What happened
The article presents a new class of continuous-time latent state-space models (LJMs) that incorporate event-driven jumps, along with a scalable training algorithm (SPVI) that accelerates inference through parallelization techniques.
Why it matters
The method addresses computational bottlenecks in training time series models for irregularly observed data, offering practical benefits for real-world applications where efficiency and accuracy are critical.
- LJMs provide an alternative to traditional ODE/SDE-based models for irregular time series
- SPVI's parallelization reduces computational costs without sacrificing accuracy
- The approach is particularly effective for systems with sparse, event-driven dynamics
- Models require fewer latent updates compared to numerical solver-based methods
The bigger picture
The SPVI method enables faster training of LJMs, making them more viable for large-scale time series analysis while maintaining accuracy comparable to traditional continuous-time models.
The paper advances time series modeling by introducing efficient, event-driven latent jump models that balance accuracy with computational practicality through novel variational inference techniques.
Technical details
- Latent jump models with Poisson-driven event processes
- Event-augmented path-space evidence lower bound (ELBO) for training
- Triangular affine scans for parallelized variational posterior sampling
- Closed-form deterministic dynamics between jumps
- State-dependent intensity and mark distributions for event modeling
FAQ
What are latent jump models (LJMs)?
LJMs are continuous-time latent state-space models that use deterministic dynamics between Poisson-driven events, with learned instantaneous jumps at event times to model irregular time series.
How does SPVI improve training efficiency?
SPVI enables parallelized variational inference by sampling complete latent event configurations before evaluating state recurrence, reducing computational depth from linear to logarithmic.
What are the key advantages of this approach?
The method achieves competitive forecast accuracy with significantly faster training (1-2 orders of magnitude) while requiring fewer latent updates than traditional numerical solvers.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by arXiv - stat.ML on Oct 12, 2026.
Comments
Sign in to join the discussion.