<ul><li>A new unsupervised learning method called recognition-parametrized Gaussian state space model (RP-GSSM) has been introduced for time series data with latent dynamical structure.</li><li>RP-GSSM is a probabilistic model that learns Markovian Gaussian latents to explain statistical dependence between observations at different time steps, combining contrastive methods with probabilistic generative models.</li><li>Unlike contrastive approaches, RP-GSSM is learned via maximum likelihood, and unlike generative approaches, it does not need an explicit network mapping from latents to observations.</li><li>The RP-GSSM allows for exact inference, maintains expressivity with a nonlinear neural network link between observations and latents, and demonstrates superior performance on tasks like learning nonlinear stochastic dynamics from video.</li></ul>

Maximum Likelihood Learning of Latent Dynamics Without Reconstruction

Discover more