![[2602.23050] Latent Matters: Learning Deep State-Space Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.23050] Latent Matters: Learning Deep State-Space Models
Summary
The paper presents a novel constrained optimization framework for training deep state-space models (DSSMs), introducing the extended Kalman VAE (EKVAE) to enhance prediction accuracy and system identification in dynamic systems.
Why It Matters
This research addresses limitations in existing DSSMs by proposing a new training approach that improves the accuracy of temporal predictions. The findings are significant for fields relying on accurate modeling of dynamic systems, such as robotics and data science.
Key Takeaways
- Introduces a constrained optimization framework for training DSSMs.
- Presents the extended Kalman VAE (EKVAE) for improved accuracy.
- Demonstrates significant enhancements in system identification.
- EKVAE outperforms traditional RNN-based DSSMs in prediction tasks.
- Successfully disentangles static and dynamic features in state-space representations.
Computer Science > Machine Learning arXiv:2602.23050 (cs) [Submitted on 26 Feb 2026] Title:Latent Matters: Learning Deep State-Space Models Authors:Alexej Klushyn, Richard Kurle, Maximilian Soelch, Botond Cseke, Patrick van der Smagt View a PDF of the paper titled Latent Matters: Learning Deep State-Space Models, by Alexej Klushyn and 4 other authors View PDF Abstract:Deep state-space models (DSSMs) enable temporal predictions by learning the underlying dynamics of observed sequence data. They are often trained by maximising the evidence lower bound. However, as we show, this does not ensure the model actually learns the underlying dynamics. We therefore propose a constrained optimisation framework as a general approach for training DSSMs. Building upon this, we introduce the extended Kalman VAE (EKVAE), which combines amortised variational inference with classic Bayesian filtering/smoothing to model dynamics more accurately than RNN-based DSSMs. Our results show that the constrained optimisation framework significantly improves system identification and prediction accuracy on the example of established state-of-the-art DSSMs. The EKVAE outperforms previous models w.r.t. prediction accuracy, achieves remarkable results in identifying dynamical systems, and can furthermore successfully learn state-space representations where static and dynamic features are disentangled. Comments: Subjects: Machine Learning (cs.LG) Cite as: arXiv:2602.23050 [cs.LG] (or arXiv:2602.23050v1 [...
