Conceptual
Browse Instruct Curate

Normalizing-Flows-Based Particle Filtering for Latent Dynamics from Image Observations

How to estimate the hidden state of a system that is only observed through images, by pairing a conditional normalizing flow (which supplies an exact observation likelihood) with jointly-learned latent linear dynamics matrices, then feeding that likelihood into a particle filter. The student learns how a fully differentiable state-space model can be trained end-to-end by maximum likelihood and used for filtering from high-dimensional observations.