Search papers, labs, and topics across Lattice.
This paper introduces Amortized Filtering and Smoothing with Flows (AFSF), a novel framework for Bayesian filtering and smoothing in high-dimensional nonlinear dynamical systems. AFSF uses a recurrent encoder to map observation histories to a fixed-dimensional summary statistic, which then conditions both a forward normalizing flow (for the filtering distribution) and a backward normalizing flow (for the backward transition kernel). By amortizing inference and sharing information between filtering and smoothing, AFSF achieves accurate approximations and supports extrapolation beyond the training horizon.
Forget Kalman filters: AFSF leverages conditional normalizing flows to amortize Bayesian filtering and smoothing in high-dimensional nonlinear systems, enabling accurate trajectory estimation and extrapolation.
Bayesian filtering and smoothing for high-dimensional nonlinear dynamical systems are fundamental yet challenging problems in many areas of science and engineering. In this work, we propose AFSF, a unified amortized framework for filtering and smoothing with conditional normalizing flows. The core idea is to encode each observation history into a fixed-dimensional summary statistic and use this shared representation to learn both a forward flow for the filtering distribution and a backward flow for the backward transition kernel. Specifically, a recurrent encoder maps each observation history to a fixed-dimensional summary statistic whose dimension does not depend on the length of the time series. Conditioned on this shared summary statistic, the forward flow approximates the filtering distribution, while the backward flow approximates the backward transition kernel. The smoothing distribution over an entire trajectory is then recovered by combining the terminal filtering distribution with the learned backward flow through the standard backward recursion. By learning the underlying temporal evolution structure, AFSF also supports extrapolation beyond the training horizon. Moreover, by coupling the two flows through shared summary statistics, AFSF induces an implicit regularization across latent state trajectories and improves trajectory-level smoothing. In addition, we develop a flow-based particle filtering variant that provides an alternative filtering procedure and enables ESS-based diagnostics when explicit model factors are available. Numerical experiments demonstrate that AFSF provides accurate approximations of both filtering distributions and smoothing paths.
