Collapsed Amortized Variational Inference for Switching Nonlinear Dynamical Systems
This addresses the problem of efficient inference in complex dynamical systems for researchers and practitioners in machine learning and time series analysis, representing an incremental improvement.
The paper tackles efficient inference in switching nonlinear dynamical systems by learning an inference network for continuous latent variables with exact marginalization of discrete ones, enabling reparameterization and end-to-end training, and demonstrates successful segmentation of time series data like videos and 3D human pose into meaningful regimes using piece-wise nonlinear dynamics.
We propose an efficient inference method for switching nonlinear dynamical systems. The key idea is to learn an inference network which can be used as a proposal distribution for the continuous latent variables, while performing exact marginalization of the discrete latent variables. This allows us to use the reparameterization trick, and apply end-to-end training with stochastic gradient descent. We show that the proposed method can successfully segment time series data, including videos and 3D human pose, into meaningful ``regimes'' by using the piece-wise nonlinear dynamics.