acceptodds
Under review as a conference paper at ICLR 2027

Outlier-Robust Variational State Estimation of Model-free Process

Abstract

We develop robust variational state estimation (RoVaS), a data-driven method for estimating the state of a model-free dynamical process from noisy measurements with outliers. We assume no knowledge of the dynamics of the process; hence, we do not have access to the state-transition model. The setting is challenging because outlier detection and dynamics learning are coupled: identifying corrupted measurements requires a model of the dynamics, while learning the dynamics requires knowing which measurements to trust. RoVaS introduces latent corruption indicators alongside the states and infers them jointly under a variational framework. A corruption gate, parameterized to saturate as the measurement magnitude grows, attenuates suspected outliers before they reach the state posterior; this yields a provable outlier-robustness guarantee at inference. On stochastic Lorenz-63 systems and real human motion-capture data, RoVaS degrades gracefully under corruption rates up to 25% and is competitive with a state-transition-model-informed robust filter, despite having no access to the state-transition model.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
State estimation, variational state estimation, dynamical variational auto encoders, outlier robustness

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QFh3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the question of estimating a dynamical system in which (1) the dynamics are unknown, (2) the measurement model on clean data is known, (3) some unknown fraction of the data is arbitrarily corrupted and carries no information about the underlying state. In the absence of a model for the dynamics, it is challenging to identify the corruptions, and prior work in this literature either had to assume a model for the dynamics or were not robust. This paper proposes the following variational method: define indicator variables for the corrupted steps with some manually chosen prior over these variables, formulate an autoregressive parametrization for the unknown dynamics as well as a specific parametrization for the posterior on the state given the measurements and the corruption indicators. For the latter, this involves a "corruption posterior", which they take to be a GRU approximating the posterior on the corruption variables conditioned on the measurements, and a "state posterior", which they take to be a Gaussian approximating the posterior on the next state conditioned on the previous trajectory and the measurements of the full trajectory apart from the next state. They jointly train these components by minimizing an ELBO loss. Under certain regularity assumptions, for instance coercivity of the network for the corruption posterior, they show that the KL between the model conditioned on corrupted data vs conditioned on clean data is finite.

The main empirical findings are that on synthetic data from a stochastic Lorenz-63 system, their method degrades the most gracefully in the presence of corruptions under linear measurements, and under nonlinear measurements it outperforms other model-free methods though it underperforms relative to a model-based baseline. For human motion capture data, it beats other baselines by a significant margin.

Critical strengths
  • The paper studies a fundamental and challenging problem in state estimation and proposes a method that is empirically shown to be more robust than other model-free methods for the same problem.
  • The ablations show that having the state posterior at time tt be independent of the measurement at time tt is indeed crucial to the performance of their method.
Critical weaknesses
  • The theoretical result, which gives a non-quantitative bound under strong assumptions that feels orthogonal to why the model is actually effective empirically, feels a bit like proving a theorem for the sake of proving a theorem and could be safely removed without affecting the overall paper.
  • The authors transparently acknowledge various limitations like knowing the measurement model for clean data and assuming independent corruptions across steps. The former feels more in scope for follow-up work , but the latter feels like a more significant shortcoming -- it is relatively unrealistic yet is the basis for the experiments as well, which exclusively involve independent corruptions across steps. While it makes sense to focus on the independent setting in the development of the methods, the paper would be strengthened by augmenting the evaluations with some examples of bursty / correlated corruptions, just to see how robust the method is.
Questions for authors

How does the method perform in settings in which the corruptions are correlated, either across time or with the state itself?

1 comment

We thank the reviewer for the constructive review and feedback.

W1. We thank the reviewer for this criticism. We agree that the theoretical result is not quantitative, but achieve an asymptotic result of the method, and is the main theorem of the WoLF method (Thm. 3.2, Duran-Martin et al., 2024). They show, and argue, that this theorem is what makes the WoLF method (in its various settings) asymptotically outlier-robust. And thus, we argue that our main proposition is what makes us call RoVaS outlier-robust, as the proposition shows that it is (asymptotically) outlier-robust. We hope this gives some insight into our thought process on the addition of the proposition. For an estimation of the KL in the proposition, we refer to Fig. 4, where we show RoVaS achieving magnitudes lower KL than VaS and VaS-tt.

W2. We thank the reviewer. We have performed correlated outlier experiments. Please see our answer to Q1.

Questions

Q1. We thank the reviewer for the question. To study outlier robustness of RoVaS and the baselines on correlated corruptions, we simulate Lorenz-63 using identity measurement map and bursty outliers using a Markov chain of mean time length 55, expected outlier probability of 1010\\%, SMNR 2020 dB, and the same hyperparameters as in the manuscript. The results are shown below.

RoVaSVaS-ttVaS
Independent corruptions (manuscript)−29.02±0.13-29.02\pm0.13−28.13±0.11-28.13\pm0.11−14.42±0.47-14.42\pm0.47
Bursts, trained on independent corruptions−26.67±0.30-26.67\pm0.30−25.45±0.07-25.45\pm0.07−15.23±0.55-15.23\pm0.55
Bursts, trained on bursts−28.23±0.31-28.23\pm0.31−25.57±0.78-25.57\pm0.78−16.44±0.83-16.44\pm0.83

The results show that both RoVaS and VaS-tt seem to handle correlated corruptions (bursty corruptions), and the generalization of training on independent outliers and testing on bursty outliers shows reasonable performance, albeit with a slightly larger drop in performance.

Reviewer_KSCiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RoVaS, an amortized variational smoother for estimating latent states when the state-transition model is unknown and measurements contain outliers. It jointly learns a dynamics prior, state posterior, and corruption probabilities. A learned corruption gate combines temporal information with a Mahalanobis-distance term to attenuate suspected outliers. The authors establish bounded posterior influence under architectural assumptions and evaluate RoVaS on stochastic Lorenz-63 and human-motion data.

Critical strengths
  1. The paper addresses a relevant problem: robust state estimation without known transition dynamics or state labels.
  2. Jointly learning dynamics, corruption indicators, and state estimates within a variational framework is an interesting approach.
  3. The reported improvements over non-robust model-free estimators are promising, particularly on Lorenz-63.
  4. The theoretical analysis provides a useful architectural guarantee for extreme outlier attenuation.
  5. The paper includes substantial implementation details, sensitivity analyses, and ablations.
Critical weaknesses
  1. The manuscript does not convincingly establish which real-world applications simultaneously require unknown transition dynamics, known measurement likelihoods, offline smoothing, and access to sufficient unlabeled training sequences. Applications such as tracking, navigation, or sensor fusion could motivate the work, but the paper should explain which assumptions hold and why existing robust smoothers or learned-dynamics estimators are inadequate. The human-motion experiment uses real trajectories but synthetically generated noise and corruptions, so it is not yet evidence of robustness to naturally occurring measurement failures.

  2. RoVaS uses complete measurement sequences, including future observations, whereas EKF and WoLF-IMQ are causal filters; the differences in temporal information are confounded with robustness and model knowledge. A more informative comparison would include model-informed robust smoothers, particularly Student-t smoothing [1], an appropriately constructed GNC-based batch smoother [2], and EMORF/S [4]. EMORF-II [5]. These references should be evaluated under matched observations, corruption realizations, and covariance assumptions. An oracle-mask smoother and a RoVaS variant using the true dynamics would help separate detection error, dynamics learning, and smoothing performance.

  3. The proposed corruption mechanism is not adequately isolated: table 6 shows that setting the Mahalanobis coefficient to zero or removing the gated input leaves NMSE essentially unchanged. In contrast, excluding the current measurement from the corresponding state factor accounts for a substantial 15.6 dB improvement; this creates a gap between the central theoretical mechanism and the empirical gains.

  4. The variational corruption update is insufficiently justified. Equation (8) combines a freely learned temporal score with a pooled Mahalanobis term, rather than deriving the corruption probabilities from the likelihood ratio implied by the generative model. Why is this preferable to a structured variational update or a semi-amortized approximation? Also, equation (3) conditions the state posterior on latent corruption indicators, whereas the implementation conditions on deterministic probabilities rather than sampled indicators. This is a restricted variational construction.

  5. The pooled mean and covariance used in the gate are estimated from measurements that include outliers, potentially biasing the very statistics used for detection. Also, the corruption-generating distribution in the Lorenz experiments has the same pooled-Gaussian form as the assumed surrogate likelihood, which makes the setting relatively favourable. For BMLMoVi, the authors change the corruption-likelihood covariance after the pooled version causes degenerate training for some seeds. The paper should examine sensitivity to contaminated statistics, misspecified corruption distributions, and train-test mismatch.

  6. The experiments primarily consider independently occurring, whole-vector measurement replacements at corruption rates up to 25%, with isotropic nominal measurement noise. Real sensor failures can be component-wise, temporally clustered, correlated across sensors, or partially informative about the state. Robustness should be evaluated under independent and correlated measurement noise, retaining matched marginal variances, and under both whole-vector and partial corruptions. In particular, comparison with EMORF/S [4] and related robust estimation methods [2, 3, 5] would be informative under correlated covariance.

  7. The theoretical guarantee needs clearer interpretation: proposition 1 proves bounded posterior influence for fixed trained parameters under sufficient assumptions, but does not quantify the bound or guarantee robustness during training. Appendix C shows that the implemented GRU-based VaS and VaS-t also exhibit bounded empirical influence because their encoders saturate, although at substantially larger magnitudes. A comparison across saturating and non-saturating encoders, together with quantitative influence bounds, would make this contribution clearer.

  8. RoVaS is motivated by amortized inference, yet no systematic runtime, memory, training-cost, or scaling comparison is presented. The experiments use state dimensions of 3 and 63 and sequence length 100. A comparison against non-amortized or semi-amortized variational inference would help establish the accuracy-efficiency tradeoff.

Ref.:

[1] R. Piché, S. Särkkä, and J. Hartikainen, “Recursive outlier-robust filtering and smoothing for nonlinear systems using the multivariate Student-t distribution,” in Proc. 2012 IEEE Int. Workshop Mach. Learn. Signal Process. (MLSP), Santander, Spain, Sep. 2012, pp. 1–6, doi: 10.1109/MLSP.2012.6349794.

[2] P. Antonante, V. Tzoumas, H. Yang, and L. Carlone, “Outlier-robust estimation: Hardness, minimally tuned algorithms, and applications,” IEEE Trans. Robot., vol. 38, no. 1, pp. 281–301, Feb. 2022, doi: 10.1109/TRO.2021.3094984.

[3] A. H. Chughtai, M. Tahir, and M. Uppal, “Bayesian heuristics for robust spatial perception,” IEEE Trans. Instrum. Meas., vol. 73, pp. 1–11, 2024, doi: 10.1109/TIM.2024.3364269.

[4] A. H. Chughtai, M. Tahir, and M. Uppal, “EMORF/S: EM-based outlier-robust filtering and smoothing with correlated measurement noise,” IEEE Trans. Signal Process., vol. 72, pp. 4318–4331, 2024, doi: 10.1109/TSP.2024.3460176.

Questions for authors
  1. Which concrete applications satisfy the joint assumptions of unknown dynamics, contaminated measurements, access to complete trajectories, and sufficient unlabeled training sequences, and what practical failure mode does RoVaS address in each application?
  2. Why does the VaS conditional KL in Figure 4 not vanish at the perturbed timestep when that measurement is excluded from the corresponding factor?
  3. Does the general boundedness proposition also establish bounded posterior influence for the implemented GRU-based VaS and VaS-t models?
  4. Why was Eq. 8 chosen instead of a corruption update based on the expected likelihood ratio?
  5. How sensitive is the gate to contaminated estimates of the pooled mean and covariance?
  6. How does the method compare under independent and correlated measurement noise when either the complete measurement vector or only a subset of components is corrupted?
  7. How does performance change when the training and test corruption rates or corruption distributions differ?
  8. What is the measured outlier-to-nominal-noise power ratio in each experiment?
  9. Which process-noise covariance was used in the Lorenz experiments?
  10. How do runtime, memory, NMSE or ELBO, and the amortization gap scale separately with state dimension, measurement dimension, and sequence length under both independent and correlated measurement noise, and after how many test sequences does amortized inference recover its training cost?
  11. Will the authors provide complete executable code for RoVaS and all matched baselines, including the correlated and diagonal implementations of EMORF/S, the EMORF-II-inspired smoother, and the GNC-cost smoother?

Minor and Reproducibility Concerns:

. Appendix D.1 appears to use process-noise covariance 0.1I, whereas Appendix D.12 states 0.01I. Correct or explain the discrepancy.

. Clarify the early-stopping procedure because the paper also reports 3,000 epochs and use of the final checkpoint.

. Specify whether a separate model is trained for every corruption rate and SMNR.

. Report the number of Monte Carlo samples used during training and inference.

. Provide full hyperparameters and tuning procedures for VaS-t, WoLF, DANSE, DNS, and VSE.

. Explain how the pooled covariance is summarized as approximately 80 times an isotropic noise covariance.

. Describe the human-motion experiment as semi-synthetic because the trajectories are real but the noise and corruption are simulated.

. Report posterior calibration, interval coverage, or predictive log scores in addition to NMSE.

4 comments

We thank the reviewer for the review and the thoughtful questions.

W1. We thank the reviewer. In signal processing, control, machine learning and allied fields, we often use models in the spirit that "All models are wrong, but some are useful". Development of useful models is crucial. There are many real-life scenarios where a-priori knowledge of the system is limited and development of useful model is difficult. In our context, it can happen that the underlying dynamical process is unknown or a model is so complex that is not easily suitable for state estimation using traditional methods like Kalman filter. For example, modeling particle movements and fluids flow in biological cells, organs, body is extremely complex. Developing suitable models for brain dynamics is still unknown, but we can collect a lot of measurements using varying type of sensing methods, such as EEG, ECoG, fMRI, MEG, etc. Then, for surveillance and defense, where modeling dynamics of various type of UAVs and drones, and their swarms is an extremely complex problem. A concrete and contemporary example is the Integrated Sensing Communication (ISAC) in 6G development, a critical problem is constructing spatio-temporal perception map where it is complex to model many moving and flying objects. In ISAC, we can collect a lot of measurements using base stations (and if required using deployed radars, lidars, etc), and thus the measurement map is well-known, but subject to occlusions and outliers (misclassified objects). Please check the website on GSM Association (GSMA): https://www.gsma.com/solutions-and-impact/industries/smart-mobility/isac-drone-detection/

W2. We thank the reviewer for the suggestion to compare with an oracle model-free smoother as well as a robust model-based smoother, to study detection error and improvement areas of RoVaS.

To this end, we implement the smoother version of the EKF, the extended Rauch-Tung-Striebel Smoother (ERTSS) but apply the WoLF generalized Bayesian updation step. We call this method WoLF-ERTSS. This means we can exploit the hyperparameters set for WoLF-IMQ for the WoLF-ERTSS. Further, we also compare with an oracle RoVaS (known outliers). The experiment to the oracle is provided as answer to reviewer QZH1 W1; the experiment of the WoLF-ERTSS is provided as answer to Q11.

W3. We thank the reviewer. Please note that the presence of coercivity term and the gated measurement input are necessary for the development of theory in the asymptotic sense of robustness to outliers. The theoretical development is in the same spirit of the paper "Outlier-robust Kalman filtering through generalised Bayes, ICML, 2024", but for a model-free dynamical process.

W4. We thank the reviewer. We proposed a corruption model in Eqn. (8) and developed the method based on dynamical variational learning, and shown its efficacy through experiments. The corruption model is based on a neural network and can thus suitably approximate any continuous function with any desired degree of accuracy; the update of this neural network is via the principled evidence lower bound (ELBO) to the log-likelihood of the evidence. We believe we have sufficiently justified the approach and principles in the article.

W5. Thank you for the comment. We agree that outliers can influence nearby state estimations (as is shown in Fig. 4 right). Please see W2 for addressing this influence. We agree that corruptions drawn from the pooled Gaussian are a favourable setting, and we will state this more clearly in the manuscript. The Mahalanobis term acts as a base corruption score, while gθg_{\boldsymbol\theta} detects corruptions from temporal context, and Proposition 1 does not require a correct pcorrp_{\mathrm{corr}}, as the reviewer points out. The change of pcorrp_{\mathrm{corr}} for BMLmovi is reported in Appendix D.10. For mismatch in the corruption rate, Table 7, 8, 9 ablate different choices of ε\varepsilon, β0\beta_0, and kk. Therein we show that training is stable across different magnitudes of these hyperparameters. For train/test mismatch, please see our answer to Q7.

W6. We thank the reviewer for the comments. To the best of the authors' knowledge, our proposed RoVaS method is the first method as an outlier-robust state estimation method for a model-free dynamical process without any a-priori knowledge of the state-transition-model of the process. The most important past work in the literature is the development of outlier-robust WoLF-IMQ method that knows the state-transition-model of the process. The paper is: "Outlier-robust Kalman filtering through generalised Bayes, International Conference on Machine Learning (ICML), 2024, authored by Gerardo Duran-Martin, Matias Altamirano, Alex Shestopaloff, Leandro Sánchez-Betancourt, Jeremias Knoblauch, Matt Jones, Francois-Xavier Briol, and Kevin Patrick Murphy". Due to knowledge of state-transition-model, WoLF-IMQ method could use traditional state estimation methods such as Kalman filter, extended Kalman filter, etc.

We performed the experiments using independent outliers, the same noise model used in the ICML 2024 paper "Outlier-robust Kalman filtering through generalised Bayes". Our main contribution lies in developing outlier-robust state estimation method for model-free process. To approach this comment, we now perform new experiments for a noise model with temporally correlated outliers. This is referred to as bursty outlier, and the experimental results are provided as an answer to the question of reviewer QFh3.

W7. We thank the reviewer for pointing this out, and we will make it clearer that the proposition is an asymptotic outlier-robust result.

W8. We thank the reviewer for suggesting alternative variational approaches, and we believe it is a very interesting future work.

Questions

Q1. Please see our response to W1.

Q2. The reviewer is correct, and we are grateful for this careful observation. The clean and corrupted posteriors were sampled with separate state histories, which is not the per-step factor of the PIF. With the same history, the per-step KL of VaS and VaS-tt at the corrupted time step is zero, since the tt'th factor omits yt\mathbf{y}_t, and we have updated the figure. The results are qualitatively the same; RoVaS shows the smallest per-step KL, and VaS-tt and VaS are significantly worse, with VaS having the poorest KL.

Q3. Yes, for GRU-based models.

Q4. We thank the reviewer for the question and suggestion. Eq. (8) was chosen so that the universal approximation theorem of neural networks is applicable (see W4), thus not restricting the corruption indicator, and can optimize the corruption indicator in our ELBO formulation. We agree that the expected likelihood ratio is another approach to pursue.

Q5. We thank the reviewer for the question. It is not very sensitive; please see Appendix F.2.

Q6. We thank the reviewer for the question. Please see our response to W6.

Q7. Please find results in the table below, comparing different train and test corruption rates of the Lorenz-63 using identity measurement maps at 20 dB SMNR.

MethodTrain \ Test0%5%10%15%20%25%
RoVaS5%−29.71-29.71−29.61-29.61−29.27-29.27−28.82-28.82−28.49-28.49−27.81-27.81
RoVaS15%−29.30-29.30−29.32-29.32−29.10-29.10−28.96-28.96−28.81-28.81−28.58-28.58
RoVaS25%−29.17-29.17−29.23-29.23−29.04-29.04−28.96-28.96−28.80-28.80−28.71-28.71
VaS-tt5%−28.90-28.90−28.38-28.38−27.82-27.82−27.06-27.06−26.39-26.39−25.38-25.38
VaS-tt15%−28.34-28.34−28.12-28.12−28.02-28.02−27.60-27.60−27.18-27.18−26.82-26.82
VaS-tt25%−27.82-27.82−27.67-27.67−27.68-27.68−27.39-27.39−27.02-27.02−26.86-26.86

We find two things interesting. First, both RoVaS and VaS-tt seem to generalize well over different corruption rates between train and test. Second, RoVaS generalizes better than VaS-tt.

Q8. Please find the table below.

ExperimentOutlier-to-noise power ratio
Lorenz63, identity, 10 dB8 (9 dB)
Lorenz63, identity, 20 dB72 (19 dB)
Lorenz63, camera, 20 dB139 (21 dB)
BMLmovi, 20 dB790 (29 dB)

Q9. We use a process noise variance of 0.1.

Q10. We thank the reviewer for the comment. Amortized inference is a single forward pass, about 0.30.3 ms per sequence in a batch of 6464 on one GPU. Training takes around 22 hours to 30003000 epochs. We agree that an amortized scaling study with measurement dimension, state dimension, training data size would be nice, but at this point we do not know a principled way to perform the study either theoretically or experimentally.

Q11. Upon acceptance of the article, code will be released for RoVaS and the baseline methods including model-based WoLF-ERTSS smoother which is the smoother version of the WoLF-IMQ method of the ICML 2024 paper "Outlier-robust Kalman filtering through generalised Bayes". Please check the comparison between RoVaS and WoLF-ERTSS in the following table, which will be included in the updated article. The results are for stochastic Lorenz-63 system with linear measurement model.

RateWoLF-ERTSS (10 dB)RoVaS (10 dB)WoLF-ERTSS (20 dB)RoVaS (20 dB)
0%−27.15±0.06-27.15 \pm 0.06−24.48±0.14-24.48 \pm 0.14−31.47±0.01-31.47 \pm 0.01−29.99±0.09-29.99 \pm 0.09
5%−26.90±0.04-26.90 \pm 0.04−24.12±0.25-24.12 \pm 0.25−31.29±0.01-31.29 \pm 0.01−29.54±0.20-29.54 \pm 0.20
10%−26.35±0.23-26.35 \pm 0.23−23.70±0.20-23.70 \pm 0.20−31.07±0.06-31.07 \pm 0.06−29.02±0.13-29.02 \pm 0.13
15%−25.89±0.15-25.89 \pm 0.15−23.14±0.28-23.14 \pm 0.28−30.80±0.09-30.80 \pm 0.09−28.87±0.15-28.87 \pm 0.15
20%−25.35±0.05-25.35 \pm 0.05−22.56±0.45-22.56 \pm 0.45−30.55±0.04-30.55 \pm 0.04−28.68±0.11-28.68 \pm 0.11
25%−24.74±0.14-24.74 \pm 0.14−22.17±0.31-22.17 \pm 0.31−30.21±0.11-30.21 \pm 0.11−28.57±0.12-28.57 \pm 0.12

We find that the difference between RoVaS and WoLF-ERTSS is around 2-3 dB on 1010 dB SMNR, and around 1.51.5-22 dB on 2020 dB SMNR. Thus, the model-free RoVaS achieves close performance to a model-based baseline.

Then we continue with the 2D camera measurement model with SMNR = 20 dB, and the results are in the following table.

RateWoLF-ERTSSRoVaS
0%−38.63±0.06-38.63 \pm 0.06−34.97±0.27-34.97 \pm 0.27
5%−38.04±0.05-38.04 \pm 0.05−34.18±0.41-34.18 \pm 0.41
10%−37.33±0.11-37.33 \pm 0.11−33.53±0.34-33.53 \pm 0.34
15%−36.57±0.07-36.57 \pm 0.07−32.76±0.44-32.76 \pm 0.44
20%−35.94±0.08-35.94 \pm 0.08−32.60±0.20-32.60 \pm 0.20
25%−34.93±0.13-34.93 \pm 0.13−32.02±0.64-32.02 \pm 0.64

For the camera measurement model, we see slightly larger differences, 3 dB to 4 dB improvement with a known STM. These results show that there are still improvements that can be made for model-free outlier-robust smoothers, which invites interesting future research.

Minor and Reproducibility Concerns: We thank the reviewer for noticing these, and we will update the manuscript accordingly.

I have gone through the authors' response: the discussion of possible applications, correction of Figure 4, oracle comparison, train–test corruption-rate results, outlier-to-noise power ratios, and clarification of the process-noise variance are useful. Nevertheless, my principal concerns remain unresolved.

  1. The rebuttal appears to clarify that the intended scope is model-free smoothing under independent measurement noise and independently generated outliers. Please state this explicitly in the manuscript; even then it does not remove the need for stronger comparisons within that setting. GNC, Student-(t), EMORF/S with a diagonalized covariance, and an EMORF-II-inspired smoother remain useful robust-smoothing benchmarks under independent measurement noise. Although these approaches are model-based, they are still informative reference methods, particularly in the synthetic experiments where the true state-transition model is available.

  2. The bursty-outlier experiment is a potentially useful stress test due to its data-generating process introduces temporal clustering in the corruption events. However, the rebuttal does not indicate that RoVaS itself has been extended with an explicit temporal model for the outlier indicators, such as a Markov or persistence prior. Also, the experiment tests empirical robustness to clustered outliers or misspecification rather than demonstrating explicit modeling of temporal outlier dynamics.

  3. The discrepancy between the ablation results and the claimed empirical importance of the corruption gate is still unresolved in author's response. Establishing that the gate and coercivity conditions are required for the asymptotic theory does not explain why removing major gate components produces almost no degradation in the reported experiments.

  4. The requested scalability analysis with respect to dimensions and sequence length remains absent.

Thus, although the rebuttal provides some useful corrections and additional results, it does not resolve the central experimental and methodological concerns as explained above; I therefore retain my original rating.

Reviewer_QZH1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces RoVaS, a method for estimating a system’s hidden states from noisy measurements that contain outliers, without a known state-transition model or clean training targets.

The main idea is to jointly estimate the states and identify corruption, then reduce the influence of suspect corruptions. A three-stage training process keeps the model from simply labeling everything as an outlier. The authors also prove that, under stated assumptions, a single arbitrarily large outlier has a bounded effect on the inferred state distribution after training.

The authors test RoVaS on simulated Lorenz-63 systems and real human motion data, with artificially added noise affecting up to 25% of measurements. RoVaS performs better than the tested model-free baselines and competes with WoLF, which knows the dynamics, although WoLF does better on the nonlinear camera test. Outlier-detection results and ablations provide further support. The experiments cover random replacement outliers, broader types of corruption remain untested.

Critical strengths
  1. The paper provides a principled way for an important and challenging problem of outlier-robust state estimation without a known state-transition model.
  2. The paper provides comparisons with several existing methods and includes ablation studies. The empirical gains are substantial across simulated dynamics and human motion data.
  3. The paper is well-written, clearly organized, and easy to follow.
Critical weaknesses
  1. Limited novelty. The proposed method mainly combines existing techniques, including variational inference, corruption indicators, and measurement gating. The necessity of joint inference over simpler alternatives is not sufficiently justified. For example, a straightforward alternative would be to first learn temporal dynamics, identify unreliable measurements using predictive residuals, and then perform state estimation with the detected outliers removed or masked. The paper does not compare against such a two-stage approach, making it difficult to assess whether the proposed joint inference framework provides a substantial improvement over simpler alternatives.

  2. Potential outlier propagation. Although the current measurement is excluded from its corresponding state posterior, the gate processes the entire raw observation sequence. Therefore, outliers may still indirectly influence other state estimates through the inference network.

  3. Limited empirical validation. The experiments mainly consider independently replaced, state-independent outliers. More importantly, the ablation study shows that removing the coercivity term or the gated measurement input has almost no effect on estimation performance in the evaluated setting. This raises questions about the practical relevance of the proposed theoretical mechanism.

Questions for authors
  1. Why is joint inference preferable to a two-stage approach that first detects and removes outliers, followed by state estimation?

  2. Can an outlier at time tt affect the state estimates at other time steps through the bidirectional gating network? How significant is this effect?

  3. How accurately does the proposed gate actually reject outliers? Could the authors report precision, recall, and false-positive rates in addition to AUROC?

  4. Given the ablation results, what practical advantages do the coercivity term and gated input provide beyond the theoretical guarantee?

  5. How robust is the method to temporally correlated outliers, partial measurement corruption, and genuine abrupt changes in the underlying dynamics?

2 comments

We thank the reviewer for the review and the thoughtful questions, and for finding the manuscript clear and well-organized.

W1. We thank the reviewer for the comments. We first address the concern of limited novelty. To the best of the authors' knowledge, our proposed RoVaS method is the first method as an outlier-robust state estimation method for a model-free dynamical process without any a priori knowledge of the state-transition-model (STM) of the process. One of the relevant past works in the literature is the development of outlier-robust WoLF-IMQ method that knows the state-transition-model of the process. The paper is: "Outlier-robust Kalman filtering through generalised Bayes, International Conference on Machine Learning (ICML), 2024, authored by Gerardo Duran-Martin, Matias Altamirano, Alex Shestopaloff, Leandro Sánchez-Betancourt, Jeremias Knoblauch, Matt Jones, Francois-Xavier Briol, and Kevin Patrick Murphy". Due to knowledge of state-transition-model, WoLF-IMQ method could use traditional state estimation methods such as Kalman filter, the extended Kalman filter, etc.

In the absence of a state-transition-model, our proposed RoVaS method can not use the traditional state estimation methods. Instead, it is necessary to develop a novel method, which is the RoVaS. In the development of RoVaS, we derive and show several aspects of outlier model-free state estimation: latent modeling of outliers and incorporation into state estimation Eq. (3)-(7), showing model-free baselines (DANSE, VSE, DNS) fail on outlier state-estimation task and showing naive optimization of Eq. (7) also fails, and we design a three stage training to mitigate 'catastrophic' failure (all measurements are either outliers or not).

Then we address the second comment, suggesting to compare with a simple baseline approach. In verbatim, the reviewer suggested "a straightforward alternative would be to first learn temporal dynamics, identify unreliable measurements using predictive residuals, and then perform state estimation with the detected outliers removed or masked". We implemented such a two-stage method using dynamical variational autoencoder (DVAE). In the first-stage of the two-stage approach, we learn the underlying state-transition-model (STM) of the process. The STM is parameterized by a neural network. Using the learned STM, we predict the residuals in the two-stage approach. In the second-stage, when a predictive (normalized) residual for a measurement is large, we consider it to be an outlier based on a threshold. The threshold is chosen carefully (using the standard logic of 9999\\% of a χ2\chi^2 distribution). This outlier prediction induces a corruption indicator mask. This corruption indicator mask is then used along with the corresponding measurement sequence in the RoVaS method to provide outlier-robust state estimation performance. The performance of the two-stage method is compared with the proposed RoVaS and an oracle method that knows the outliers exactly. We experiment on stochastic Lorenz-63 system with identity measurements using SMNR at 2020 dB (same hyperparameters as in the manuscript). The NMSE performance versus percentage of corruption outliers (pϵp_\epsilon) is shown in the following Table. From the Table, we observe that the two-stage approach suffers significantly for larger corruption rates. We will include this result in the updated manuscript.

Method0%5%10%15%20%25%
Two-stage (VaS STM)−29.33±0.07-29.33 \pm 0.07−29.30±0.19-29.30 \pm 0.19−28.82±0.06-28.82 \pm 0.06−27.88±0.15-27.88 \pm 0.15−21.71±0.80-21.71 \pm 0.80−20.19±1.11-20.19 \pm 1.11
RoVaS−29.97±0.06-29.97 \pm 0.06−29.41±0.17-29.41 \pm 0.17−29.07±0.12-29.07 \pm 0.12−28.79±0.01-28.79 \pm 0.01−28.65±0.12-28.65 \pm 0.12−28.62±0.12-28.62 \pm 0.12
Oracle−29.30±0.15-29.30 \pm 0.15−29.22±0.21-29.22 \pm 0.21−29.18±0.13-29.18 \pm 0.13−29.06±0.11-29.06 \pm 0.11−28.78±0.14-28.78 \pm 0.14−28.69±0.15-28.69 \pm 0.15

W2. We thank the reviewer for this point and agree outliers can influence other time-steps. However, this is, as far as we understand, unavoidable. This is since we have no prior knowledge of what is an outlier or not, and thus, must use the measurements we have to train our model. As far as we can see, this is shared by any Bayesian state estimation method that does not know what is an outlier, including the outlier-robust WoLF-IMQ method of the paper "Outlier-robust Kalman filtering through generalised Bayes, ICML 2024".

W3. We thank the reviewer. We performed the experiments using independent outliers, the same noise model used in the ICML 2024 paper "Outlier-robust Kalman filtering through generalised Bayes". Our main contribution lies in developing outlier-robust state estimation method for model-free process. However, we now extend with new experiments for a noise model with temporally correlated outliers. We refer to this as bursty outlier, and the experimental results are provided as an answer to the question of reviewer QFh3.

Then comes the concern of ablation study and theory. Please note that the presence of coercivity term and the gated measurement input are necessary for the development of theory in the asymptotic sense of robustness to outliers. The theoretical development is in the same spirit of the paper "Outlier-robust Kalman filtering through generalised Bayes, ICML, 2024", but for a model-free dynamical process.

Questions

Q1.

Please see our response to W1.

Q2.

We thank the reviewer for this question. This investigation is done in Appendix C, see e.g. Fig. 4 (right). For VaS and VaS-tt, the effect of an outlier is significant with respect to the state estimation of neighboring time stamps; the per-step KL is in the order of 100100. For RoVaS, the per-step KL is in the order of 0.10.1. This per-step KL degrades gracefully the further away we get from an outlier (see Fig. 4 right).

Q3.

We thank the reviewer for this suggestion, and the results can be found in the table below. The results show that the gate is effective.

SMNRRateAUROCPrecisionRecallFPR (%)F1F_1
10 dB5%0.921±0.0010.921 \pm 0.0010.978±0.0120.978 \pm 0.0120.473±0.0650.473 \pm 0.0650.05±0.020.05 \pm 0.020.635±0.0630.635 \pm 0.063
10 dB10%0.928±0.0040.928 \pm 0.0040.993±0.0030.993 \pm 0.0030.537±0.0020.537 \pm 0.0020.04±0.020.04 \pm 0.020.697±0.0010.697 \pm 0.001
10 dB15%0.916±0.0230.916 \pm 0.0230.996±0.0020.996 \pm 0.0020.551±0.0100.551 \pm 0.0100.04±0.020.04 \pm 0.020.710±0.0080.710 \pm 0.008
10 dB20%0.921±0.0030.921 \pm 0.0030.994±0.0010.994 \pm 0.0010.532±0.0120.532 \pm 0.0120.08±0.020.08 \pm 0.020.693±0.0100.693 \pm 0.010
10 dB25%0.906±0.0160.906 \pm 0.0160.998±0.0010.998 \pm 0.0010.531±0.0030.531 \pm 0.0030.03±0.010.03 \pm 0.010.693±0.0030.693 \pm 0.003
20 dB5%0.959±0.0090.959 \pm 0.0090.987±0.0140.987 \pm 0.0140.909±0.0140.909 \pm 0.0140.06±0.070.06 \pm 0.070.946±0.0070.946 \pm 0.007
20 dB10%0.994±0.0020.994 \pm 0.0020.998±0.0010.998 \pm 0.0010.947±0.0030.947 \pm 0.0030.02±0.010.02 \pm 0.010.971±0.0020.971 \pm 0.002
20 dB15%0.988±0.0090.988 \pm 0.0090.995±0.0040.995 \pm 0.0040.952±0.0080.952 \pm 0.0080.08±0.070.08 \pm 0.070.973±0.0050.973 \pm 0.005
20 dB20%0.994±0.0010.994 \pm 0.0010.999±0.0010.999 \pm 0.0010.946±0.0020.946 \pm 0.0020.03±0.020.03 \pm 0.020.972±0.0010.972 \pm 0.001
20 dB25%0.994±0.0010.994 \pm 0.0010.998±0.0010.998 \pm 0.0010.946±0.0040.946 \pm 0.0040.05±0.020.05 \pm 0.020.972±0.0020.972 \pm 0.002

An interesting observation, however, is the recall at SMNR 10 dB, that is around 0.5, which we interpret as around half of the outliers are predicted as outliers and the other half is not. This might not be surprising, as 10 dB is a relatively noisy environment; thus, distinguishing between outliers and actual measurements is a hard task, especially without known dynamics (model-free).

Q4.

We thank the reviewer for the question. In the ablation of the coercivity term and gated input (Table 6), the corruptions are drawn from a Gaussian fitted to the measurements, so they are of typical magnitude, and removing the coercivity term or the gated input does not change NMSE. Both components target large corruptions, ∥yt∥→∞\|\mathbf{y}_t\| \to \infty. To test this, we perturbed one measurement in each test sequence by a vector of norm CC (Lorenz-63, identity measurements, 20 dB SMNR, 10% corruption, 5 seeds, 50 sequences each; an extract of the Fig. 4 experiment):

RoVaSRoVaS, λ=0\lambda=0RoVaS, no gated input
NMSE, C=0C=0−29.12-29.12−29.20-29.20−29.17-29.17
NMSE, C=104C=10^4−29.07-29.07−28.49-28.49−22.17-22.17
Median PIF, C=105C=10^5414123223228402840

A single large corruption does not affect RoVaS. Without the gated input, NMSE degrades by 7 dB. Without the coercivity term, NMSE degrades by 0.7 dB, and the KL to the unperturbed posterior is 232 compared to RoVaS 41. Thus, both the coercivity term and the gated input help reduce the influence of very large outliers.

Q5.

We thank the reviewer for these interesting suggestions, and refer to the answer to the question of QFh3 for experiment with bursty outliers. In general, training on independent outliers and testing on bursty outliers degrades the NMSE by a bit more than 2 dB. However, training and testing on bursty outliers is more stable, and degrades the NMSE by less than 1 dB. In both cases, RoVaS shows promising performance and beats both VaS-tt and VaS.

Comments on the paper

1 comment

General comment

We thank the reviewers for the reviews and feedback. Here we summarize the new experiments and the changes; the details are in the individual responses.

To clarify the scope of the manuscript: The manuscript investigates if outlier-robust state estimation is possible for a model-free process, i.e., without knowing the dynamical model of the underlying process (without having a state-transition-model (STM) of the process). Traditional state estimation methods, such as Kalman filter, extended Kalman filter, particle filters, are STM-based. The STM-based methods compare a measurement to a prediction using the known STM and then decide whether a measurement is useful or outlier. The most relevant work of STM-based outlier-robust method is in the following paper "Outlier-robust Kalman filtering through generalised Bayes, International Conference on Machine Learning (ICML), 2024, authored by Gerardo Duran-Martin, Matias Altamirano, Alex Shestopaloff, Leandro Sánchez-Betancourt, Jeremias Knoblauch, Matt Jones, Francois-Xavier Briol, and Kevin Patrick Murphy". In the paper, the outlier-robust state estimation method is based on Kalman filter and full knowledge of process dynamics (STM).

In the model-free setting, we do not know the dynamics and do not have the STM. Therefore, in the learning process, outlier detection and dynamics learning are coupled. To the best of our knowledge, the proposed Robust Variational State Estimation (RoVaS) method is the first outlier-robust method for a model-free process. We develop the RoVaS, and provide theoretical and experimental studies. We provide three experimental settings, with two different signal-to-measurement-noise (SMNR) values, across 5 different corruption rates, to provide evidence of the efficacy of the RoVaS method.

As per the review comments, we summarize the new experiments in this rebuttal:

  • Oracle with the true corruption indicators: the RoVaS method is close to this oracle (within training variance).
  • Two-stage baseline, first learning STM, then learning state estimation: the two-stage approach degrades significantly for higher corruption rates: up to 99 dB worse NMSE compared to RoVaS.
  • Model-based (STM-informed) outlier-robust smoother: RoVaS is within 1.51.5 to 44 dB in NMSE to a model-based outlier-robust smoother across main experiments.
  • Burst corruptions: RoVaS is still superior in NMSE; 11 to 2.52.5 dB better compared to baselines.
  • Train/test mismatch in corruption rate: RoVaS generalizes well across different train and test corruption rates, and better than VaS-tt.
  • Precision, recall and false-positive rate of the corruption gate.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.