pith. machine review for the scientific record. sign in

arxiv: 2603.05011 · v2 · submitted 2026-03-05 · 📡 eess.SY · cs.SY

Receding-Horizon Maximum-Likelihood Estimation of Neural-ODE Dynamics and Thresholds from Event Cameras

Kazumune Hashimoto , Kazunobu Serizawa , Masako Kishida This is my paper

Pith reviewed 2026-05-15 15:44 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords event camerasneural ODEmaximum likelihood estimationreceding horizondynamics identificationcontrast thresholdasynchronous sensing
0
0 comments X

The pith

A receding-horizon maximum-likelihood estimator recovers Neural ODE dynamics parameters and contrast thresholds from event camera streams.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes an online method to identify continuous-time dynamics from asynchronous event data. It models the latent state with a Neural ODE, maps it to log-intensity, and treats events as a history-dependent point process whose intensity depends on an unknown contrast threshold. The approach uses a receding window of recent events, performs a few gradient steps on the joint likelihood per update, and approximates the compensator integral for efficiency. Synthetic tests show that dynamics parameters and the threshold can be recovered together, with explicit trade-offs between estimation accuracy and update latency as the window length varies.

Core claim

The receding-horizon maximum-likelihood estimator jointly recovers the Neural ODE parameters and the contrast threshold by optimizing the event likelihood over sliding data windows, using a differentiable state-to-image mapping and a smooth surrogate for the contrast-threshold trigger, while storing only two scalars per pixel for streaming operation.

What carries the argument

Receding-horizon maximum-likelihood optimization over a sliding window of events, with the log-likelihood formed from a marked point process whose conditional intensity is a smooth surrogate of contrast-threshold crossing.

If this is right

  • Joint recovery of dynamics and threshold parameters is feasible from synthetic event streams.
  • Longer horizon windows improve accuracy at the cost of increased estimation latency.
  • Streaming computation is achieved by retaining only last-event time and log-intensity per pixel.
  • Monte Carlo pixel subsampling keeps the compensator integral tractable during online updates.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same estimator structure could be adapted to other asynchronous sensors that produce threshold-crossing events.
  • Real-time parameter tracking might enable adaptive control loops that update the dynamics model on the fly.
  • Extensions could incorporate uncertainty quantification on the recovered parameters for safety-critical applications.

Load-bearing premise

The Neural ODE together with its smooth surrogate intensity function must accurately represent the true continuous dynamics and the physical event-triggering process.

What would settle it

Run the estimator on real event data from a physical system whose true dynamics and contrast threshold are known independently; mismatch between recovered and true values would falsify the modeling assumption.

Figures

Figures reproduced from arXiv: 2603.05011 by Kazumune Hashimoto, Kazunobu Serizawa, Masako Kishida.

Figure 1
Figure 1. Figure 1: Illustration of the proposed framework. naturally accommodates irregular timestamps and history de￾pendence. A practical challenge is computational cost. Point￾process likelihoods include, in addition to a sum over observed events, a normalization (no-event) term that integrates the predicted event rate over time and sums over the full pixel grid. Evaluating this term can dominate runtime especially for lo… view at source ↗
Figure 2
Figure 2. Figure 2: Surrogate intensity (11) as a function of the distance [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: True pixel dependent threshold map C(·). TABLE I: Estimated parameters and key hyperparameters in the numerical experiments (pixel-dependent threshold only). Item Value / description Renderer / data (fixed) Grid size (Himg, Wimg) 64 × 64 Base center cbase = (32, 32) Drift parameters A = B = 22 px, T1 = 1.0s, T2 = 1.3s Intensity params (Ibg, Iamp, σ) = (0.15, 0.75, 2.0) Log eps (data / model) ϵ = 10−3 Frame… view at source ↗
Figure 4
Figure 4. Figure 4: Representative snapshots from the synthetic event-camera sequence. Top: rendered intensity frames. Bottom: corre [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Learning curves for a fixed horizon H = 15. Top: αˆ versus episode e. Bottom: ωˆ versus episode e. Dashed lines indicate the ground truth. where A ⊆ Ω denotes the set of active pixels (pixels that emitted at least one event in the sequence), so that the threshold error is not dominated by unobserved regions. Figs. 7 and 8 report the RMSE values as functions of H. The estimation of ω is particularly sensiti… view at source ↗
Figure 6
Figure 6. Figure 6: Snapshots of pixel-dependent contrast-threshold estimation over online episodes ( [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Mean update time per step as a function of [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Horizon ablation. Threshold-map error RMSEC (H) over active pixels (38) as a function of H. Note that the y￾axis is shown in linear scale. [3] C. Brandli, R. Berner, M.-H. Yang, S.-C. Liu, and T. Delbruck, “A 240×180 130 db 3 µs latency global shutter spatiotemporal vision sensor,” IEEE Journal of Solid-State Circuits, vol. 49, no. 10, pp. 2333– 2341, 2014. [4] A. Z. Zhu, L. Yuan, K. Chaney, and K. Daniili… view at source ↗
read the original abstract

Event cameras emit asynchronous brightness-change events where each pixel triggers an event when the last event exceeds a threshold, yielding a history-dependent measurement model. We address online maximum-likelihood identification of continuous-time dynamics from such streams. The latent state follows a Neural ODE and is mapped to predicted log-intensity through a differentiable state-to-image model. We model events with a history-dependent marked point process whose conditional intensity is a smooth surrogate of contrast-threshold triggering, treating the contrast threshold as an unknown parameter. The resulting log-likelihood consists of an event term and a compensator integral. We propose a receding-horizon estimator that performs a few gradient steps per update on a receding horizon window. For streaming evaluation, we store two scalars per pixel (last-event time and estimated log-intensity at that time) and approximate the compensator via Monte Carlo pixel subsampling. Synthetic experiments demonstrate joint recovery of dynamics parameters and the contrast threshold, and characterize accuracy--latency trade-offs with respect to the window length.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a receding-horizon maximum-likelihood estimator for jointly identifying Neural-ODE parameters and the contrast threshold from asynchronous event-camera streams. The latent state evolves according to a Neural ODE, is mapped to log-intensity via a differentiable state-to-image model, and events are modeled as a history-dependent marked point process whose intensity is a smooth surrogate of threshold triggering; the log-likelihood (event term plus compensator integral) is maximized by a few gradient steps on each receding window, with Monte Carlo pixel subsampling used to approximate the compensator for streaming operation. Synthetic experiments are presented to demonstrate joint recovery and accuracy-latency trade-offs versus window length.

Significance. If the synthetic recovery results prove robust under higher-fidelity integration and quantitative benchmarking, the framework would supply a statistically principled online method for continuous-time system identification from event data, directly addressing a gap in event-based vision for control and robotics by treating the sensor threshold as an estimable parameter rather than a fixed constant.

major comments (2)
  1. [§4 (Synthetic Experiments)] §4 (Synthetic Experiments): the central claim of joint recovery of Neural-ODE weights and contrast threshold is supported only by qualitative or unreported quantitative results; the abstract and available description provide no parameter-error metrics, baselines, number of Monte Carlo trials, or error bars, leaving the strength of the evidence for the main contribution unclear.
  2. [Compensator approximation] Compensator approximation (streaming implementation): Monte Carlo subsampling of pixels to approximate the compensator integral stores only two scalars per pixel and performs only a few gradient steps per window; the resulting stochastic gradient variance directly affects updates to both the Neural-ODE parameters and the threshold, yet no analysis or variance-reduction technique is provided to bound its effect on recovery accuracy.
minor comments (2)
  1. [Abstract] Abstract: the description of the synthetic data-generation process, network architecture, and exact form of the smooth surrogate intensity function is omitted, making it difficult to reproduce or assess the modeling assumptions.
  2. [Notation] Notation: the precise functional dependence of the surrogate intensity on the contrast threshold parameter should be stated explicitly (e.g., as an equation) rather than described only in prose.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which have helped us strengthen the presentation of our results. We address each major comment below and have revised the manuscript to incorporate additional quantitative evidence and analysis.

read point-by-point responses

  1. Referee: [§4 (Synthetic Experiments)] the central claim of joint recovery of Neural-ODE weights and contrast threshold is supported only by qualitative or unreported quantitative results; the abstract and available description provide no parameter-error metrics, baselines, number of Monte Carlo trials, or error bars, leaving the strength of the evidence for the main contribution unclear.

    Authors: We agree that explicit quantitative metrics strengthen the evidence. In the revised manuscript we have added a new table in §4 reporting mean parameter recovery error (MSE on Neural-ODE weights and on the contrast threshold) across 20 independent Monte Carlo trials, together with standard-error bars for each window length. We also include a baseline comparison against a fixed-threshold estimator that does not jointly optimize the contrast parameter. These additions directly quantify the joint-recovery performance claimed in the abstract. revision: yes

  2. Referee: Compensator approximation (streaming implementation): Monte Carlo subsampling of pixels to approximate the compensator integral stores only two scalars per pixel and performs only a few gradient steps per window; the resulting stochastic gradient variance directly affects updates to both the Neural-ODE parameters and the threshold, yet no analysis or variance-reduction technique is provided to bound its effect on recovery accuracy.

    Authors: We acknowledge that an explicit variance analysis was missing. The revised §3.3 now includes a short derivation of the Monte Carlo variance of the compensator estimator (proportional to 1/M where M is the number of subsampled pixels) and reports an empirical study showing that the gradient variance remains below 0.015 for M=128 across the tested window lengths. We also note that the receding-horizon scheme with only a few gradient steps per window limits error accumulation; a full theoretical convergence bound under stochastic gradients is left for future work as it would require additional assumptions on the Neural-ODE Lipschitz constants. revision: partial

Circularity Check

0 steps flagged

No significant circularity; derivation follows independent MLE principles

full rationale

The paper constructs the log-likelihood directly from the marked point-process model with conditional intensity defined via the Neural-ODE state and smooth surrogate, independent of the fitted parameters. The receding-horizon estimator applies standard gradient steps to this likelihood on a sliding window, while the Monte Carlo pixel subsampling is an implementation approximation for the compensator integral rather than a definitional reduction. No equations reduce the claimed recovery of dynamics and threshold to the inputs by construction, no uniqueness theorems or ansatzes are imported via self-citation, and the synthetic experiments function as external validation rather than fitted-input predictions. The central claim therefore remains self-contained against the modeling assumptions.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard assumptions from neural ODEs and point-process modeling plus two fitted parameter classes; no new entities are postulated.

free parameters (2)
  • Neural ODE network weights
    Parameters of the neural network that defines the vector field of the continuous-time dynamics are optimized via gradient steps on the likelihood.
  • contrast threshold
    Scalar parameter treated as unknown and recovered jointly with the dynamics.
axioms (2)
  • domain assumption The state-to-image mapping is differentiable
    Required to compute gradients of the log-likelihood with respect to the Neural ODE parameters.
  • domain assumption The smooth surrogate conditional intensity correctly models the marked point process of events
    The likelihood is constructed from this surrogate; if it deviates from true triggering statistics the recovered parameters lose physical meaning.

pith-pipeline@v0.9.0 · 5479 in / 1377 out tokens · 49047 ms · 2026-05-15T15:44:41.836616+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 46 canonical work pages

  1. [1]

    Event-based vision: A survey,

    G. Gallegoet al., “Event-based vision: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 1, pp. 154–180, 2022

    work page 2022

  2. [2]

    A 128×128 120 db 15µs latency asynchronous temporal contrast vision sensor,

    P. Lichtsteiner, C. Posch, and T. Delbruck, “A 128×128 120 db 15µs latency asynchronous temporal contrast vision sensor,”IEEE Journal of Solid-State Circuits, vol. 43, no. 2, pp. 566–576, 2008. 9 Fig. 6: Snapshots of pixel-dependent contrast-threshold estimation over online episodes (H= 15).Left:ground-truth threshold fieldC(u)used for event synthesis.Rig...

    work page 2008

  3. [3]

    A 240×180 130 db 3µs latency global shutter spatiotemporal vision sensor,

    C. Brandli, R. Berner, M.-H. Yang, S.-C. Liu, and T. Delbruck, “A 240×180 130 db 3µs latency global shutter spatiotemporal vision sensor,”IEEE Journal of Solid-State Circuits, vol. 49, no. 10, pp. 2333– 2341, 2014

    work page 2014

  4. [4]

    Ev-flownet: Self- supervised optical flow estimation for event-based cameras,

    A. Z. Zhu, L. Yuan, K. Chaney, and K. Daniilidis, “Ev-flownet: Self- supervised optical flow estimation for event-based cameras,” inRobotics: Science and Systems (RSS), 2018. Fig. 9: Mean update time per step as a function ofH

    work page 2018

  5. [5]

    E-RAFT: Dense optical flow from event cameras,

    M. Gehrig, M. Millh ¨ausler, D. Gehrig, and D. Scaramuzza, “E-RAFT: Dense optical flow from event cameras,” inInternational Conference on 3D Vision (3DV). IEEE, 2021, pp. 197–206

    work page 2021

  6. [6]

    Dense continuous-time optical flow from event cameras,

    M. Gehrig, M. Muglikar, and D. Scaramuzza, “Dense continuous-time optical flow from event cameras,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 7, pp. 4736–4746, 2024

    work page 2024

  7. [7]

    Events-to-video: Bringing modern computer vision to event cameras,

    H. Rebecq, R. Ranftl, V . Koltun, and D. Scaramuzza, “Events-to-video: Bringing modern computer vision to event cameras,” inProc. IEEE/CVF CVPR, 2019

    work page 2019

  8. [8]

    Continuous-time intensity estimation using event cameras,

    C. Scheerlinck, N. Barnes, and R. Mahony, “Continuous-time intensity estimation using event cameras,” inAsian Conference on Computer Vision (ACCV), 2018, pp. 308–324

    work page 2018

  9. [9]

    High frame rate video reconstruction based on an event camera,

    L. Pan, R. Hartley, C. Scheerlinck, M. Liu, X. Yu, and Y . Dai, “High frame rate video reconstruction based on an event camera,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 5, pp. 2519–2533, 2022

    work page 2022

  10. [10]

    Time lens: Event-based video frame interpolation,

    S. Tulyakov, D. Gehrig, S. Georgoulis, J. Erbach, M. Gehrig, Y . Li, and D. Scaramuzza, “Time lens: Event-based video frame interpolation,” in Proc. IEEE/CVF CVPR, 2021, pp. 16 155–16 164

    work page 2021

  11. [11]

    The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam,

    E. Mueggler, H. Rebecq, G. Gallego, T. Delbruck, and D. Scaramuzza, “The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam,”The International Journal of Robotics Research, vol. 36, no. 2, pp. 142–149, 2017

    work page 2017

  12. [12]

    Esim: an open event camera simulator,

    H. Rebecq, D. Gehrig, and D. Scaramuzza, “Esim: an open event camera simulator,” inConference on Robot Learning (CoRL), 2018

    work page 2018

  13. [13]

    Event-driven sensing for efficient perception: Vision and audition algorithms,

    S.-C. Liu, B. Rueckauer, E. Ceolini, A. E. G. Huber, and T. Delbruck, “Event-driven sensing for efficient perception: Vision and audition algorithms,”IEEE Signal Processing Magazine, vol. 36, no. 6, pp. 29– 37, 2019

    work page 2019

  14. [14]

    Learning adaptive parameter representation for event-based video reconstruction,

    D. Gu, J. Li, and L. Zhu, “Learning adaptive parameter representation for event-based video reconstruction,”IEEE Signal Processing Letters, vol. 31, pp. 1950–1954, 2024. 10

    work page 1950

  15. [15]

    Event-based stereo depth es- timation by temporal-spatial context learning,

    W. Chen, Y . Zhang, X. Sun, and F. Wu, “Event-based stereo depth es- timation by temporal-spatial context learning,”IEEE Signal Processing Letters, vol. 31, pp. 1429–1433, 2024

    work page 2024

  16. [16]

    Perfect recovery and sensitivity analysis of time encoded bandlimited signals,

    A. A. Lazar and L. T. T ´oth, “Perfect recovery and sensitivity analysis of time encoded bandlimited signals,”IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 51, no. 10, pp. 2060–2073, 2004

    work page 2060

  17. [17]

    Regularized signal reconstruction for level-crossing sampling using slepian functions,

    S. Senay, J. Oh, and L. F. Chaparro, “Regularized signal reconstruction for level-crossing sampling using slepian functions,”Signal Processing, vol. 92, no. 4, pp. 1157–1165, 2012

    work page 2012

  18. [18]

    Asynchronous process- ing of sparse signals,

    A. Can-Cimino, E. Sejdi ´c, and L. F. Chaparro, “Asynchronous process- ing of sparse signals,”IET Signal Processing, vol. 8, no. 3, pp. 257–266, 2014

    work page 2014

  19. [19]

    Time encoding via unlimited sampling: Theory, algorithms and hardware validation,

    D. Florescu and A. Bhandari, “Time encoding via unlimited sampling: Theory, algorithms and hardware validation,”IEEE Transactions on Signal Processing, vol. 70, pp. 4912–4924, 2022

    work page 2022

  20. [20]

    Robust online reconstruction of continuous-time signals from a lean spike train ensemble code,

    A. Chattopadhyay and A. Banerjee, “Robust online reconstruction of continuous-time signals from a lean spike train ensemble code,”IEEE Transactions on Signal Processing, vol. 73, pp. 2008–2021, 2025

    work page 2008

  21. [21]

    Filtering and detection for doubly stochastic Poisson processes,

    D. L. Snyder, “Filtering and detection for doubly stochastic Poisson processes,”IEEE Transactions on Information Theory, vol. 18, no. 1, pp. 91–102, 1972

    work page 1972

  22. [22]

    Estimating a state-space model from point process observations,

    A. C. Smith and E. N. Brown, “Estimating a state-space model from point process observations,”Neural Computation, vol. 15, no. 5, pp. 965–991, 2003

    work page 2003

  23. [23]

    Dy- namic analysis of neural encoding by point process adaptive filtering,

    U. T. Eden, L. M. Frank, R. Barbieri, V . Solo, and E. N. Brown, “Dy- namic analysis of neural encoding by point process adaptive filtering,” Neural Computation, vol. 16, no. 5, pp. 971–998, 2004

    work page 2004

  24. [24]

    Event-triggered proximal online gradient de- scent algorithm for parameter estimation,

    Y . Zhou and G. Chen, “Event-triggered proximal online gradient de- scent algorithm for parameter estimation,”IEEE Transactions on Signal Processing, vol. 72, pp. 2594–2606, 2024

    work page 2024

  25. [25]

    Event-triggered state estimation through confidence level,

    W. Liu, “Event-triggered state estimation through confidence level,” IEEE Transactions on Signal Processing, vol. 73, pp. 1337–1350, 2025

    work page 2025

  26. [26]

    Asymptotic error rates for point process classifi- cation,

    X. Rong and V . Solo, “Asymptotic error rates for point process classifi- cation,”IEEE Transactions on Signal Processing, vol. 73, pp. 738–750, 2025

    work page 2025

  27. [27]

    A bayesian mixture model of temporal point processes with determinantal point process prior,

    Y . Dong, S. Ye, Q. Han, Y . Cao, H. Xu, and H. Yang, “A bayesian mixture model of temporal point processes with determinantal point process prior,”IEEE Transactions on Signal Processing, vol. 73, pp. 2216–2226, 2025

    work page 2025

  28. [28]

    Reducing the sim-to-real gap for event cameras,

    T. Stoffregen, C. Scheerlinck, D. Scaramuzza, T. Drummond, N. Barnes, L. Kleeman, and R. Mahony, “Reducing the sim-to-real gap for event cameras,” inComputer Vision – ECCV 2020, ser. Lecture Notes in Computer Science, vol. 12372. Springer, 2020, pp. 534–549

    work page 2020

  29. [29]

    Event camera calibration of per-pixel biased contrast threshold,

    Z. Wang, Y . Ng, P. van Goor, and R. Mahony, “Event camera calibration of per-pixel biased contrast threshold,” inAustralasian Conference on Robotics and Automation (ACRA), 2019, author-accepted version available as arXiv:2012.09378

    work page arXiv 2019

  30. [30]

    State space models for event cameras,

    N. Zubi ´c, M. Gehrig, and D. Scaramuzza, “State space models for event cameras,” inProc. IEEE/CVF CVPR, 2024, pp. 5819–5828

    work page 2024

  31. [31]

    A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation,

    G. Gallego, H. Rebecq, and D. Scaramuzza, “A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation,” inProc. IEEE CVPR, 2018

    work page 2018

  32. [32]

    Event cameras, contrast maximization and reward functions: an analysis,

    T. Stoffregen and L. Kleeman, “Event cameras, contrast maximization and reward functions: an analysis,” inProc. IEEE/CVF CVPR, 2019

    work page 2019

  33. [33]

    The spatio-temporal poisson point process: A simple model for the alignment of event camera data,

    C. Gu, E. Learned-Miller, D. Sheldon, G. Gallego, and P. Bideau, “The spatio-temporal poisson point process: A simple model for the alignment of event camera data,” inProc. IEEE/CVF ICCV, 2021

    work page 2021

  34. [34]

    Constrained linear state estimation—a moving horizon approach,

    C. V . Rao, J. B. Rawlings, and J. H. Lee, “Constrained linear state estimation—a moving horizon approach,”Automatica, vol. 37, no. 10, pp. 1619–1628, 2001

    work page 2001

  35. [35]

    Constrained state estimation for nonlinear discrete-time systems: Stability and moving horizon approximations,

    C. V . Rao, J. B. Rawlings, and D. Q. Mayne, “Constrained state estimation for nonlinear discrete-time systems: Stability and moving horizon approximations,”IEEE Transactions on Automatic Control, vol. 48, no. 2, pp. 246–258, 2003

    work page 2003

  36. [36]

    Moving horizon estimation for discrete-time linear systems with binary sensors: Algorithms and stability results,

    G. Battistelli, L. Chisci, and S. Gherardini, “Moving horizon estimation for discrete-time linear systems with binary sensors: Algorithms and stability results,”Automatica, vol. 85, pp. 374–385, 2017

    work page 2017

  37. [37]

    Event-triggered maximum likelihood state estimation,

    D. Shi, T. Chen, and L. Shi, “Event-triggered maximum likelihood state estimation,”Automatica, vol. 50, no. 1, pp. 247–254, 2014

    work page 2014

  38. [38]

    Moving horizon estimation meets multi-sensor information fusion: Development, opportunities and challenges,

    L. Zou, Z. Wang, J. Hu, and Q.-L. Han, “Moving horizon estimation meets multi-sensor information fusion: Development, opportunities and challenges,”Information Fusion, vol. 60, pp. 1–10, 2020

    work page 2020

  39. [39]

    Neu- ral ordinary differential equations,

    R. T. Q. Chen, Y . Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neu- ral ordinary differential equations,” inAdvances in Neural Information Processing Systems, vol. 31, 2018, pp. 6571–6583

    work page 2018

  40. [40]

    Nonlinear filtering with counting observations,

    A. Segall, M. H. A. Davis, and T. Kailath, “Nonlinear filtering with counting observations,”IEEE Transactions on Information Theory, vol. 21, no. 2, pp. 143–149, 1975

    work page 1975

  41. [41]

    Online variational inference for state-space models with point-process observations,

    A. Zammit-Mangion, K. Yuan, V . Kadirkamanathan, M. Niranjan, and G. Sanguinetti, “Online variational inference for state-space models with point-process observations,”Neural Computation, vol. 23, no. 8, pp. 1967–1999, 2011

    work page 1967

  42. [42]

    Recurrent marked temporal point processes: Embedding event history to vector,

    N. Du, H. Dai, R. Trivedi, U. Upadhyay, M. Gomez-Rodriguez, and L. Song, “Recurrent marked temporal point processes: Embedding event history to vector,” inProc. ACM SIGKDD (KDD), 2016, pp. 1555–1564

    work page 2016

  43. [43]

    The neural hawkes process: A neurally self-modulating multivariate point process,

    H. Mei and J. M. Eisner, “The neural hawkes process: A neurally self-modulating multivariate point process,” inAdvances in Neural Information Processing Systems, 2017

    work page 2017

  44. [44]

    Latent ordinary differential equations for irregularly-sampled time series,

    Y . Rubanova, R. T. Q. Chen, and D. Duvenaud, “Latent ordinary differential equations for irregularly-sampled time series,” inAdvances in Neural Information Processing Systems, 2019

    work page 2019

  45. [45]

    Neural jump ordinary differ- ential equations: Consistent continuous-time prediction and filtering,

    C. Herrera, F. Krach, and J. Teichmann, “Neural jump ordinary differ- ential equations: Consistent continuous-time prediction and filtering,” in International Conference on Learning Representations (ICLR), 2021

    work page 2021

  46. [46]

    D. J. Daley and D. Vere-Jones,An Introduction to the Theory of Point Processes: Volume II: General Theory and Structure, 2nd ed. Springer, 2008

    work page 2008