differentiable particle filters are an emerging class of sequential Bayesian
inference techniques that use neural networks to construct components in state
space models. Existing approaches are mostly based on offline supervised
training strategies. This leads to the delay of the model