Deep Spectral Learning of Embedded Latent Transfer Operators for Stochastic Dynamical Systems
Researchers have developed a novel spectral learning method called Deep Spectral Encoder (DSE) for analyzing stochastic nonlinear dynamical systems. DSE utilizes an operator-based latent state-space model where a neural encoder maps observations to Markovian latent states. The method employs functional canonical correlation analysis to derive state coordinates, from which transfer and observation operators are estimated. This approach allows for generalized sequential Bayesian filtering and Koopman spectral mode decomposition, demonstrating superior performance over existing baselines in noisy and partially observable scenarios. AI