Alexej Klushyn

Forget RNNs for modeling dynamics: a Kalman filter-VAE hybrid can learn state-space representations that disentangle static and dynamic features, outperforming previous models in prediction accuracy and system identification.