Reviews: A Disentangled Recognition and Nonlinear Dynamics Model for Unsupervised Learning

The paper presents a time-series model for high dimensional data by combining variational auto-encoder (VAE) with linear Gaussian state space model (LGSSM). The proposed model takes the latent repressentation from VAE as the output of LGSSM. The exact inference of linear Gaussian state space model via Kalman smoothing enables efficient and accurate variational inference for the overall model. To extend the temporal dynamics beyond linear dependency, the authors use a LSTM to parameterize the matrices in LGSSM. The performance of the proposed model is evaluated through bouncing ball and Pendulum experiments.