Differentiable Particle Filtering

• Differentiable Particle Filtering is a family of sequential Monte Carlo methods that integrates neural network parameterizations to enable gradient-based learning through its sampling, weighting, and resampling steps.
• It leverages techniques like reparameterization, soft-resampling, and optimal transport to overcome the inherent non-differentiability of classical particle filters.
• This approach has shown improved accuracy and robustness in complex, non-linear, and high-dimensional systems by directly optimizing dynamic, observation, and proposal models.

Differentiable Particle Filtering describes a family of sequential Monte Carlo (SMC) algorithms in which the classical particle filtering procedure—sampling, weighting, and resampling over latent state sequences in state-space models—is rendered differentiable with respect to model and proposal parameters, typically via neural network parameterizations and algorithmic modifications. By enabling backpropagation through the filtering process, such approaches facilitate direct optimization of parameterized dynamic, observation, and proposal models in complex, non-linear, and high-dimensional state spaces, especially when model components are implicit or data-driven functions (e.g., deep neural networks) (Chen et al., 2023).

1. Foundations and State-Space Model Structure

At the core of differentiable particle filtering is the discrete-time nonlinear state-space model,

$$p(x_{1:T},o_{1:T}\mid a_{1:T}) = p(x_1) \prod_{t=2}^T p(x_t\mid x_{t-1},a_t) \prod_{t=1}^T p(o_t\mid x_t)$$

where $x_t$ denotes the latent state (possibly high-dimensional), $a_t$ the control/action, and $o_t$ the observation. The objective is online estimation of the filtering posterior $p(x_t\mid o_{1:t},a_{1:t})$ given only observed signals and potentially learned models for state transition and observation likelihood (Wan et al., 21 Jul 2025, Chen et al., 2021).

Classic SMC methods represent the filtering posterior by a set of weighted samples (particles), but these are inherently non-differentiable due to discrete sampling and resampling. Differentiable particle filtering (DPF) frameworks replace or circumvent these obstacles, allowing the filtering operation to be optimized end-to-end using stochastic gradient descent.

2. Differentiability Mechanisms in Particle Filtering

Achieving differentiability in PF poses two main algorithmic hurdles:

• Propagation (Sampling): Standard sampling steps are discrete and non-differentiable. DPFs utilize the reparameterization trick: particles are generated via a deterministic function of previous particles, observations, and fixed base noise, i.e., $x_t^i = f_\theta(x_{t-1}^i,a_t,\epsilon_t^i)$ with $\epsilon_t^i\sim\mathcal N(0,I)$, so gradients can flow through $\theta$ (Jonschkowski et al., 2018, Chen et al., 2021).
• Resampling: Classical resampling schemes—multinomial, stratified—are non-differentiable. DPFs deploy various strategies:
  • Soft-resampling: Adds a uniform component to the weight vector, yielding a mixture between resampling and random copying; differentiable weight-correction factors restore unbiasedness (Dupty et al., 2024, Chen et al., 2023).
  • Optimal-transport resampling: Applies an entropy-regularized Sinkhorn algorithm to calculate a smooth, differentiable map from weighted particles to resampled particles (Corenflos et al., 2021).
  • Stop-gradient surrogates: In lieu of modifying the forward pass, backpropagation is altered with stop-gradient (SG) operations in resampling, matching the Fisher identity's score estimator (Ścibior et al., 2021, Brady et al., 29 Oct 2025).
  • Neural or learned resamplers: Transformer-based attention modules compute barycentric mapping of particles in a fully differentiable way (Dupty et al., 2024, Chen et al., 2024).

For multi-modal or highly non-Gaussian posteriors, DPFs can employ expressive proposal distributions, such as conditional normalizing flows or learned generative samplers (e.g., diffusion models) able to capture the complexity of the target posterior (Wan et al., 21 Jul 2025, Chen et al., 2021).

3. Model, Proposal, and Likelihood Parameterization

Differentiable particle filters frequently parameterize key components as neural networks to enable flexible adaptation from raw data:

• Dynamics/Transition Model ($p(x_t \mid x_{t-1}, a_t)$): Implemented as MLPs, RNNs, or invertible flows, often with reparameterized noise injection to allow gradient-based updates (Li et al., 2023, Chen et al., 2024).
• Observation/Likelihood Model ($p(o_t \mid x_t)$): CNNs, GNNs, or learned metric spaces in the observation domain (including vision-based tasks) (Jonschkowski et al., 2018, Nierop et al., 10 Jan 2025).
• Proposal/Importance Density ($x_t$0): Modern DPFs replace the classic bootstrap proposal with parameterized, observation-guided flows, Gaussian mixtures, or deep generative models (e.g., conditional diffusion or normalizing flow structures) (Chen et al., 2021, Cox et al., 2024, Wan et al., 21 Jul 2025).

The joint system is fully differentiable—either via reparameterization or by employing continuous relaxations/SG-tricks as described above—allowing gradient information to propagate to all parameter sets via backpropagation through time.

4. Loss Functions and Training Objectives

End-to-end objectives used to optimize DPFs include:

• (Supervised) Negative Log-Likelihood or RMSE: When true latent trajectories are known, such as $x_t$1 or RMSE between estimated and true state (Jonschkowski et al., 2018, Li et al., 2023).
• Variational Sequential Monte Carlo ELBO: The standard ELBO for particle filters,

$x_t$2

is unbiased and amenable to unsupervised and semi-supervised learning (Nierop et al., 10 Jan 2025, Li et al., 2024, Chen et al., 2023).

• Joint and Mixed Losses: In practice, semi-supervised settings mix an ELBO and a supervised (MSE) loss, or employ blockwise pseudo-likelihood surrogates when only partial labels exist (Wen et al., 2020, Li et al., 2023, Li et al., 2024).
• Score Matching Loss for Generative Samplers: For diffusion-based DPFs, a denoising score-matching loss on noise prediction is utilized (Wan et al., 21 Jul 2025).

Optimization proceeds via stochastic gradient-based methods (Adam, RMSProp), with gradients flowing through the entire particle filtering pipeline.

5. Algorithmic Variants: Advanced Conditional and Generative Proposals

Recent research has extended DPF architectures far beyond simple bootstrap models. Notable developments include:

• Conditional Normalizing Flow Proposals: Flow-based models enable arbitrarily flexible proposal densities by stacking invertible mappings conditioned on observations, yielding lower importance weight variance and better coverage of multi-modal posteriors (Chen et al., 2021, Chen et al., 2024, Cox et al., 2024).
• DiffPF and Generative Diffusion Samplers: Direct particle generation from a learned conditional diffusion model, conditioned on prior predictions and fused observation features, allows unbiased sampling from highly non-Gaussian and multi-modal distributions without the need for importance weighting or resampling (Wan et al., 21 Jul 2025). This framework achieves equal-weighted particle clouds and shows state-of-the-art empirical accuracy for challenging vision-based and multi-modal localization tasks.
• Online and Regime-Switching DPFs: On-the-fly learning methods update model and proposal parameters using only the ELBO from streaming data (no labels required) (Li et al., 2023). Regime-learning DPFs learn both switching structure and per-regime behavior in systems with discontinuous or switching dynamics (e.g., interacting multiple model frameworks) (Brady et al., 2024, Brady et al., 2024).
• Graph and Structured Data: Differentiable PF algorithms have been extended to GNN frameworks, allowing for approximately universal graph representations via sequential individualized refinements modeled as GNN particle transitions and differentiable resampling (Dupty et al., 2024).

6. Empirical Results and Comparative Performance

Across a range of benchmark tasks, DPF methods demonstrate robustness and superior performance versus traditional PFs when models are misspecified, data are high-dimensional, or posteriors are multimodal/nonlinear:

• DiffPF achieves up to 82.8% accuracy improvement (RMSE) in highly multimodal maze localization and 26% lower error on KITTI odometry compared to strong baselines (Wan et al., 21 Jul 2025).
• Conditional normalizing flow DPFs substantially reduce tracking errors in vision-based disk tracking and improve effective sample sizes (Chen et al., 2021).
• OT-based resamplers yield smooth, low-variance gradient estimators and improved ELBOs in both synthetic and real-world SSM tasks (Corenflos et al., 2021, Csuzdi et al., 2024, Nierop et al., 10 Jan 2025).
• Semi-supervised DPFs can exploit extremely limited labeled data, outperforming RNN and LSTM baselines by large margins in state estimation, especially in robotics scenarios (Wen et al., 2020, Li et al., 2024).
• Flow-based and GMM proposals further reduce estimation bias and address degeneracy in nonlinear, high-dimensional SSMs (Chen et al., 2024, Cox et al., 2024).
• Online or adaptive learning approaches adapt to distribution shifts in real time, matching oracle DPFs optimized with latent state access (Li et al., 2023).

7. Limitations, Practical Considerations, and Future Directions

Despite significant advances, DPF implementations face scalability and robustness challenges:

• Computational Cost: Differentiable resampling (OT, flows, diffusion) typically scales as $x_t$3 or $x_t$4 per step and may be impractical for very large particle sets or state dimension. Runtime can be mitigated via architectural choices (shallow U-Nets, reduced particle/step count) (Wan et al., 21 Jul 2025).
• Gradient Variance: Score-function estimators (REINFORCE) used for discrete steps remain high-variance; SG-based or OT relaxations ameliorate this but may bias gradients.
• Expressivity and Stability: Extremely multimodal or high-dimensional posteriors may still challenge GMM or flow-based approximators. Continuous relaxations must balance smoothness with bias.
• Resampling Biases: Soft, neural, or transport-based resamplers bias the filtering distribution depending on regularization/gain parameters; careful tuning and theoretical guarantees are necessary (Corenflos et al., 2021, Csuzdi et al., 2024).
• Label Scarcity: Semi-supervised ELBO objectives succeed in low-dimensional or well-specified domains but may yield limited improvement for high-dimensional tasks with deep observation models unless paired with auxiliary losses (Li et al., 2024).

Key areas for future research include unbiased differentiable resamplers for discrete and continuous state spaces, scaling up DPFs to longer sequences and larger particle counts, integrating with adaptive particle allocation, and further hybridization with deep generative modeling frameworks (e.g., conditional diffusion, normalizing flows) (Wan et al., 21 Jul 2025, Nierop et al., 10 Jan 2025, Chen et al., 2021).

---

References

• (Wan et al., 21 Jul 2025) DiffPF: Differentiable Particle Filtering with Generative Sampling via Conditional Diffusion Models (2025)
• (Chen et al., 2021) Differentiable Particle Filters through Conditional Normalizing Flow (2021)
• (Corenflos et al., 2021) Differentiable Particle Filtering via Entropy-Regularized Optimal Transport (2021)
• (Li et al., 2023) Learning Differentiable Particle Filter on the Fly (2023)
• (Chen et al., 2024) Normalizing Flow-based Differentiable Particle Filters (2024)
• (Wen et al., 2020) End-To-End Semi-supervised Learning for Differentiable Particle Filters (2020)
• (Jonschkowski et al., 2018) Differentiable Particle Filters: End-to-End Learning with Algorithmic Priors (2018)
• (Dupty et al., 2024) PF-GNN: Differentiable particle filtering based approximation of universal graph representations (2024)
• (Cox et al., 2024) Learning state and proposal dynamics in state-space models using differentiable particle filters and neural networks (2024)
• (Nierop et al., 10 Jan 2025) Deep Variational Sequential Monte Carlo for High-Dimensional Observations (2025)
• (Ścibior et al., 2021) Differentiable Particle Filtering without Modifying the Forward Pass (2021)
• (Csuzdi et al., 2024) Differentiable Particle Filtering using Optimal Placement Resampling (2024)
• (Li et al., 2024) Revisiting semi-supervised training objectives for differentiable particle filters (2024)
• (Brady et al., 2024) Regime Learning for Differentiable Particle Filters (2024)
• (Brady et al., 2024) Differentiable Interacting Multiple Model Particle Filtering (2024)
• (Brady et al., 29 Oct 2025) PyDPF: A Python Package for Differentiable Particle Filtering (2025)
• (Chen et al., 2023) An overview of differentiable particle filters for data-adaptive sequential Bayesian inference (2023)