Differentiable Particle Filters through Conditional Normalizing Flow

Abstract

Differentiable particle filters provide a flexible mechanism to adaptively train dynamic and measurement models by learning from observed data. However, most existing differentiable particle filters are within the bootstrap particle filtering framework and fail to incorporate the information from latest observations to construct better proposals. In this paper, we utilize conditional normalizing flows to construct proposal distributions for differentiable particle filters, enriching the distribution families that the proposal distributions can represent. In addition, normalizing flows are incorporated in the construction of the dynamic model, resulting in a more expressive dynamic model. We demonstrate the performance of the proposed conditional normalizing flow-based differentiable particle filters in a visual tracking task.

Efficient Learning of the Parameters of Non-Linear Models using Differentiable Resampling in Particle Filters

“…A method called stop-gradient resampling is outlined in [24] which only modifies the messages used in backpropagation, is implemented in PyTorch and makes use of automatic differentiation libraries. In [25], they utilise conditional normalisation flows to construct flexible probability distributions for differentiable particle filters. A comparison of differentiable filters can be seen in [26].…”

“…Evaluate the final log likelihood, log p(y 1:T |θ), and associated derivative, d dθ log p(y 1:T |θ), using (16) and(25), respectively.…”