{"id":"arxiv/2604.07169","url":"https://arxiv.org/abs/2604.07169","tags":["stat.ML","cs.LG","math.NA"],"score":0,"title":"Amortized Filtering and Smoothing with Conditional Normalizing Flows","domain":"stat.ML","source":"arxiv","authors":["Tiangang Cui","Xiaodong Feng","Chenlong Pei","Xiaoliang Wan","Tao Zhou"],"reviews":0,"subject":"stat.ML","summary":"Bayesian filtering and smoothing are central to data assimilation in nonlinear dynamical systems. Recent advances in deep generative models provide flexible approximations of the associated non-Gaussian posterior distributions. However, several existing approaches require the score to be re-estimated or a transport map to be constructed at each assimilation step. We propose an amortized framework for filtering and smoothing that reuses trained conditional models across observation sequences and assimilation times. The framework jointly learns a shared recurrent summary network and two…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-04-08","reactions":{}}