[Paper Review] Variational Gaussian Process State-Space Models

Roger Frigola, Yutian Chen+1 more|arXiv (Cornell University)|Jun 18, 2014

Gaussian Processes and Bayesian Inference17 references99 citations

TL;DR

This paper proposes a variational Bayesian inference framework for nonlinear state-space models using sparse Gaussian processes, enabling efficient, scalable learning with uncertainty quantification. By combining variational inference with sequential Monte Carlo in an auxiliary system, the method achieves linear computational complexity and tractable posterior approximation, outperforming MCMC-based approaches in prediction speed while maintaining accuracy.

ABSTRACT

State-space models have been successfully used for more than fifty years in different areas of science and engineering. We present a procedure for efficient variational Bayesian learning of nonlinear state-space models based on sparse Gaussian processes. The result of learning is a tractable posterior over nonlinear dynamical systems. In comparison to conventional parametric models, we offer the possibility to straightforwardly trade off model capacity and computational cost whilst avoiding overfitting. Our main algorithm uses a hybrid inference approach combining variational Bayes and sequential Monte Carlo. We also present stochastic variational inference and online learning approaches for fast learning with long time series.

Motivation & Objective

• To develop a scalable, nonparametric Bayesian approach for learning complex nonlinear dynamical systems without overfitting.
• To enable tractable posterior inference over nonlinear state transition and observation functions using Gaussian process priors.
• To reduce computational cost of predictions compared to MCMC-based smoothing methods, especially for long time series.
• To support online and stochastic variational inference for real-time or large-scale learning scenarios.
• To provide a principled model for systems where prior knowledge of the dynamics is limited or unavailable.

Proposed method

• Uses sparse Gaussian processes as nonparametric priors over the state transition and observation functions to encode smoothness and continuity.
• Employs a hybrid inference strategy: variational Bayes for approximate posterior over latent states and hyperparameters, combined with sequential Monte Carlo for smoothing in an auxiliary system.
• Introduces a variational lower bound that enables efficient optimization via expectation propagation and stochastic gradient updates.
• Applies stochastic variational inference (SVI) and online learning to scale to long time series with reduced training time.
• Uses a fixed-lag particle smoother for inference in the auxiliary system, enabling efficient posterior approximation.
• Derives a tractable posterior distribution over the nonlinear dynamical system, allowing fast probabilistic predictions independent of time series length.

Experimental results

Research questions

• RQ1 Can a nonparametric Bayesian approach with Gaussian processes learn complex nonlinear state-space models without overfitting?
• RQ2 How can variational inference be combined with sequential Monte Carlo to achieve scalable learning in nonlinear SSMs?
• RQ3 Can the computational complexity of posterior prediction be made independent of time series length in GP-SSMs?
• RQ4 To what extent can stochastic and online variational inference accelerate learning on long time series?
• RQ5 How well can the model capture complex dynamical behaviors such as limit cycles in neural spike data?

Key findings

• The variational GP-SSM achieved a test RMSE of 1.15 on the Lorenz-96 dataset, outperforming GP-NARX and linear models.
• The model achieved a log predictive density of -1.61, comparable to PMCMC-based methods but with significantly faster inference (0.14s vs. 421s for test predictions).
• Stochastic variational inference (SVI) reduced training time to 4.12 minutes for a 10,000-length time series, compared to 547 minutes for PMCMC.
• The model successfully captured a 6 Hz oscillatory pattern and limit cycle dynamics in rat hippocampal spike train data, with posterior samples showing consistent attractor behavior.
• The posterior over the state transition function revealed high confidence near the limit cycle, indicating robust learning of nonlinear dynamics.
• The approach enables fast, uncertainty-aware predictions that are independent of time series length, unlike MCMC-based methods.