Latent Matters: Learning Deep State-Space Models

2026-02-26

AI Summary

This work addresses the challenge that deep state-space models often fail to accurately capture the underlying dynamics of observed sequences during training. To overcome this limitation, the authors propose a constrained optimization framework that integrates amortized variational inference with classical Bayesian filtering and smoothing techniques, resulting in an Extended Kalman Variational Autoencoder (EKVAE). By imposing structured constraints, the framework enables a disentangled representation of static and dynamic latent variables. Experimental results demonstrate that EKVAE significantly outperforms existing state-of-the-art models in system identification and long-term prediction tasks, while successfully learning semantically meaningful and disentangled latent state representations.

Technology Category

• Intelligent Robots: State Estimation
• Machine Learning: Deep Generative Models & Autoencoders
• Search and Optimization: Mixed Discrete/Continuous Search

Application Category

• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

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.

Problem

• deep state-space models
• system identification
• temporal predictions
• underlying dynamics
• evidence lower bound

Innovation

• constrained optimisation
• deep state-space models
• extended Kalman VAE
• system identification
• disentangled representation