acceptodds
Under review as a conference paper at ICLR 2027

Structured Physical Primitives for Robust Neural Decoding

Abstract

The main challenge of neural decoding in practical applications, especially in brain-computer interfaces (BCIs), lies in maintaining stable performance in the presence of noise and missing data. Current methods either learn dynamics directly from data or impose statistical constraints, making them prone to degradation when data quality declines. Physical motion, however, possesses intrinsic geometric structure with motion-specific primitives, such as translation and rotation in rodent planar motion, which are independent of neural data quality and offer a stable source of prior knowledge when neural evidence is unreliable. Motivated by this, we propose Structured Physical Primitives for Neural Decoding (MaPIC), a novel framework comprising two complementary components: a neural evidence encoder for observation likelihood and a physical primitive model for dynamic Bayesian priors. By maximizing the ELBO, MaPIC continuously integrates neural evidence with physical priors, where the KL divergence term serves as a structured regularizer guided by physical motion primitives. Notably, MaPIC learns the posterior uncertainty of each primitive via variational inference, enabling adaptive weighting among primitives. Our framework is causal and plug-and-play, requiring no modifications to the base architecture. Experiments on simulated dynamical systems and real rat hippocampal recordings demonstrate that even when of neurons is randomly dropped, MaPIC surpasses state-of-the-art methods, significantly reducing position decoding error and correcting directional phase delays at turning points.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Neural decoding, Dynamical systems, Robust inference, Computational neuroscience

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ADtaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MaPIC, a framework for robust neural decoding that integrates learned neural evidence with structured physical motion priors. The approach combines a neural encoder that models observation likelihoods with a Bayesian dynamic model incorporating physical motion primitives, such as translation and rotation. The components are integrated through variational inference, where the KL divergence encourages consistency with the physical priors, while learned posterior uncertainties enable adaptive weighting of the different primitives. The proposed framework is designed to be causal and applicable to existing decoding architectures. The approach is evaluated on simulated dynamical systems and real rat hippocampal recordings, demonstrating improved decoding accuracy and robustness to missing neural observations compared with the considered baselines.

Critical strengths
  • The central idea of incorporating explicit physical motion primitives into neural decoding is appealing. By combining learned neural evidence with structured dynamic priors, the proposed approach seeks to exploit complementary sources of information, particularly when the neural observations become unreliable.

-The paper evaluates the proposed approach on both simulated dynamical systems and real neural recordings, including scenarios where subsets of neurons are removed. The reported improvements in decoding accuracy, particularly in the presence of missing neural observations, are encouraging and suggest that the proposed approach may provide practical benefits for robust neural decoding.

Critical weaknesses
  • A central claim of the paper concerns the novelty of integrating learned neural decoding with physically motivated dynamic models. However, this general direction has been explored in previous works combining neural networks with model-based Bayesian inference and dynamical priors, e.g., "Exploring the Trade-off Between Deep-Learning and Explainable Models for Brain-Machine Interfaces" (NeurIPS 2024).

  • Although the experiments demonstrate improved performance when neural observations are artificially degraded through neuron dropout, this evaluation addresses only one aspect of the robustness challenges motivating the paper. In particular, it is not sufficiently clear what training data and supervision are required to learn the proposed model, whether it must be trained separately for each subject or recording session, and whether the learned physical priors can transfer across sessions or subjects. These questions are particularly important given the claimed practical advantages of the approach. Moreover, robustness to limited calibration data and non-stationarity is a central challenge in neural decoding, as emphasized by Karpowicz et al., "Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark" (NeurIPS 2024).

Questions for authors
  • Could the authors clarify what data and supervision are required to train the proposed framework? In particular, does learning the physical primitive model require labeled kinematic trajectories, and is the model trained separately for each subject or recording session? How is the proposed plug-and-play property realized in practice, and does applying MaPIC to a new subject or session require retraining or calibration?

  • Please elaborate on the methodological differences between MaPIC and existing approaches that combine learned neural decoding with dynamical models? In particular, what advantages arise specifically from the representation in terms of physical motion primitives and their inferred uncertainties, compared with more conventional learned or model-based Bayesian dynamics?

  • Are the neuron dropout patterns considered in the evaluation introduced exclusively at inference, or are similar dropout patterns also included during training? Moreover, to what extent are the reported improvements attributable to the structured physical primitives and their uncertainty-based adaptive weighting, rather than simply to the introduction of an additional dynamic prior?

  • Have the authors evaluated the proposed framework in cross-session or few-shot adaptation settings, where the available labeled data for a new recording session are limited? Such experiments, motivated by settings considered in the FALCON benchmark, would be particularly valuable for assessing whether the proposed physical priors indeed facilitate robust decoding under practical non-stationarity and limited supervision.

2 comments
AuthorsOpenReview ↗

Training with Frozen Decoders and the Role of Motion Correction Under Neuronal Dropout

We sincerely thank the reviewer for their helpful suggestions.

Q1: Training Data, Supervision, and the Plug-and-Play Property.

Response:

Learning the physical primitive model requires labeled kinematic trajectories, and the model is trained separately for each recording session. No labels are needed at inference. Plug-and-play is realized by keeping the base decoder frozen: MaPIC takes only the decoder's output and is trained on top of it, so the base decoder is never modified or retrained. For a new subject or session, the frozen base decoder requires no retraining for MaPIC, and only the MaPIC module is trained on that session's labeled trajectories.

The code is released and can be inspected at any time: https://anonymous.4open.science/r/MaPIC-F671/

Q2/W1: Methodological Differences from Existing Decoding–Dynamics Approaches.

Response:

The methodological difference is that existing approaches, including KalmanNet and the heteroscedastic Kalman filter of Cubillos et al. (NeurIPS 2024), adapt only the balance between neural evidence and the model's motion prediction, while MaPIC also corrects the motion prediction itself before fusion, along the physical motion primitives. This gives two advantages. First, when the dynamic model is inaccurate and neural observations are unreliable, MaPIC corrects the biased prediction that a model-based Bayesian filter must follow: error is 39–52% lower than a geometry-aware EKF under a loss of 80% of neurons (Table 5). Second, the inferred uncertainty of each primitive sets the fusion weight, and one module handles circular, linear and toroidal variables on three real datasets.

The work of Cubillos et al. is now cited, and we compare directly with KalmanNet and model-based Kalman filters on rat CA1; the row below shows Achilles, with all four rats in App. A.2, Table 6.

Achilles (cm)Frozen base decoderKalman, constant covarianceKalman, adaptive covarianceKalmanNetGRU + Kalman updateMaPIC
Position error5.77 ± 0.58 *5.88 ± 0.72 *6.28 ± 1.29 *5.96 ± 0.60 *5.96 ± 0.87 *5.12 ± 0.51

Mean ± s.d. over 9 paired runs. * p < 0.05 against MaPIC, paired Wilcoxon test.

Q3/W2: Dropout in Training and Attribution of the Improvements.

Response:

Similar dropout patterns are also included during training, and the improvements are attributable to the structured physical primitives and their uncertainty-based adaptive weighting, not to an additional dynamic prior alone. We evaluate two scenarios (App. A.1.2): persistent neuron dropout, where a random fraction of neurons is removed for the whole test session, and transient signal loss, where observations are degraded or removed within a short window. During training, random windows have random neurons dropped at random rates, with the same augmentation for all trained methods, and the test patterns are drawn independently.

On attribution (App. A.1.3–A.1.4): a filter with the same dynamic prior and the same adaptive observation covariance, but without our learned primitive correction, is matched by MaPIC when the prior is exact and is 39–52% worse when it is not. Removing the learned primitive correction roughly doubles the error under 80% dropout (13.72° to 29.46°), and replacing uncertainty-based adaptive weighting with a fixed weight significantly slows recovery after signal loss (4.56° to 5.81°); both p < 0.05 over 10 seeds.

Inaccuracy of the dynamic priorEKF with the same dynamic priorMaPIC without learned primitive correctionMaPIC with a fixed weightMaPIC
none19.38 ± 6.0021.77 ± 8.4122.11 ± 8.7016.96 ± 5.35
0.528.62 ± 9.90 *29.46 ± 9.45 *19.07 ± 5.65 *13.72 ± 2.26
1.026.36 ± 9.97 *26.75 ± 10.50 *17.55 ± 6.77 *14.30 ± 3.76
1.522.20 ± 9.54 *22.38 ± 9.72 *14.65 ± 5.2913.56 ± 3.44

Median angular error during a 2-s loss of 80% of neurons, mean ± s.d. over 10 seeds. * p < 0.05 against MaPIC, two-sided paired Wilcoxon test.

Q4: Cross-Session and Few-Shot Adaptation.

Response:

Cross-session and few-shot adaptation are not the focus of the current work, but we will address them in future work following the FALCON settings. Under limited supervision within a session, the physical priors already facilitate robust decoding: in simulation the MaPIC module is trained with two labeled trials (16 s), and on grid cells MaPIC trained on 20% of the data (9.28 cm) outperforms both baselines trained on all of it (10.16 and 9.40 cm). FALCON is now cited in Related Work.

Thank you for the detailed responses and additional experiments. The responses have adequately addressed my concerns.

Reviewer_p9EcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MaPIC, a causal temporal correction method for continuous neural decoding of behavior. It adds a correction module to a frozen base decoder, fuses neural observations and motion predictions according to their estimated uncertainties, and updates the decoded state on the behavioral state manifold. The authors evaluate MaPIC on synthetic data and several rat neural recording datasets, focusing on decoding performance with limited training data and under random neuron dropout.

Critical strengths

This paper addresses an important challenge in continuous behavior decoding: maintaining performance when neural recordings are noisy or incomplete. Combining state-space geometry with causal temporal correction offers a useful alternative to frame-by-frame decoding or generic smoothing. The evaluation covers controlled circular and toroidal state spaces and several rat neural recording datasets. In the linear-track experiments, MaPIC is applied to multiple frozen base decoders. The torus experiments also analyze errors near angular boundaries, providing targeted evidence for the role of manifold-aware correction.

Critical weaknesses

The paper does not sufficiently clarify how MaPIC differs from existing uncertainty-aware filtering and trajectory-structured decoding methods. In particular, the authors should explain the incremental contributions of its adaptive fusion and motion prior. They should also provide more details on how motion primitives, analytical motion terms, and observation covariance are specified for the real-data experiments. The statistical support for some of the paper’s conclusions is limited. The paper does not report paired significance tests or confidence intervals, and the torus experiments use only three random seeds. In addition, MaPIC-Full performs slightly worse than MaPIC-NoGate on the Achilles dataset. The authors should explain this result and qualify the related claims as appropriate.

Questions for authors
  1. How does this work distinguish MaPIC from existing uncertainty-aware filtering, trajectory-structured decoding, and methods for robustness to neural recording drift? In particular, what is MaPIC’s substantive contribution beyond the adaptive fusion already used in Kalman-based methods?
  2. For each real dataset, how are the motion primitives and any analytical motion terms defined, and how are reliability features mapped to observation covariance? Is this mapping calibrated? What is the scope of the diagonal Gaussian, first-order inertia, and conditional independence assumptions?
  3. Were paired significance tests performed for the comparisons in the paper? Given that the torus experiments use only three random seeds, could the authors provide confidence intervals or other uncertainty estimates and clarify the statistical basis for claims such as “significant improvement” and “outperforms the baselines in every condition”?
  4. On the Achilles dataset, MaPIC-Full has slightly higher mean error than MaPIC-NoGate, although the difference is small relative to the standard deviations. Is this difference consistent across runs? Could it be related to dataset characteristics or the conditions under which adaptive gating is beneficial?
1 comment
AuthorsOpenReview ↗

Learning Motion Corrections Before Fusion: New Filter Comparisons and Paired Analyses

We sincerely thank the reviewer for their helpful suggestions.

Q1: Contribution Beyond Kalman-Based Adaptive Fusion.

Response:

MaPIC's contribution beyond adaptive fusion is that it corrects the motion prediction itself before fusion, whereas Kalman-based methods only adapt the fusion weight and take the motion prediction as given. The motion prior is fused through jointly learned weights rather than fixed in advance, and the same module runs on circular, linear and toroidal state spaces. We compare directly with Kalman filters, KalmanNet and a GRU with a Kalman update on rat CA1. The row below shows Achilles and Buddy, and all four rats are in the new App. A.2, Table 6.

The code for all filter comparisons is released: https://anonymous.4open.science/r/MaPIC-F671/

Position error (cm)Frozen base decoderKalman (constant covariance)Kalman (adaptive covariance)KalmanNetGRU + Kalman updateMaPIC
Achilles5.77 ± 0.58 *5.88 ± 0.72 *6.28 ± 1.29 *5.96 ± 0.60 *5.96 ± 0.87 *5.12 ± 0.51
Buddy13.60 ± 2.74 *14.08 ± 2.75 *13.23 ± 3.0214.44 ± 2.92 *13.24 ± 3.0512.54 ± 2.32

Mean ± s.d. over 9 paired runs. * p < 0.05 against MaPIC, paired Wilcoxon test.

Q2: Specification of Motion Primitives and Observation Covariance.

Response:

The real datasets use the same construction as the synthetic experiments, with no analytical motion term, and the mapping to observation covariance is calibrated by likelihood fitting. The motion primitives are rotation on the circle for head direction, translation along the track for CA1, and translation along the two phase directions of the torus for grid cells. There is no analytical motion term on real data, and the motion prediction is learned from labelled training trajectories.

The six reliability features are readout confidence, an implausible-jump indicator, jump magnitude, spike count, the analytical motion term at the readout (zero on real data), and observation availability. The observation covariance is not learned freely: it is first fitted to the readout error on labelled data by maximum likelihood under a heteroscedastic Gaussian model, its overall scale is selected on validation, and it is then trained end to end. It is therefore a likelihood-fitted estimate of readout error variance.

The diagonal Gaussian, first-order inertia and conditional independence assumptions apply to smooth motion with unimodal errors, which covers the locomotion tasks studied here (Sec. 3.5, Discussion).

Q3: Statistical Support for the Reported Improvements.

Response:

Yes, paired significance tests are now reported.

The synthetic circle experiments use 10 seeds with two-sided paired Wilcoxon tests (Tables 3–5), and the CA1 comparison uses 9 paired runs per rat with the same test (Table 6); all are reported as mean ± s.d. The improvement is consistent at the level of individual runs: on CA1, MaPIC has lower error than the Kalman filter in all 36 paired runs and is significantly better than the frozen base decoder on all four rats, and on the circle, under mismatch and 80% neuron loss, it has lower error than the geometry-aware EKF in 9–10 of 10 seeds. The torus experiments use three seeds and are reported as mean ± s.d.; there the text now reads "lowest mean error in all 16 conditions". "Significant improvement" is used only where a test supports it.

Q4: MaPIC-NoGate on Achilles.

Response:

This difference reflects dataset characteristics rather than a consistent effect: it is 0.03 cm against a standard deviation of about 0.4 cm, so the two are equivalent on Achilles. Achilles has the most accurate base decoder of the four rats, so readout reliability varies little and adaptive gating has little to adapt to. Where the base decoder is less accurate, gating is clearly beneficial: on Cicero it reduces error from 21.24 to 17.93 cm.

Reviewer_7XrbOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces MaPIC, a temporal correction method for neural decoding that combines instantaneous neural estimates with motion priors defined on a behavioral state manifold. The method uses a GRU to infer motion variables, regularizes them with a Gaussian autoregressive prior, and combines motion predictions with neural observations using uncertainty-dependent weights. They perform experiments on synthetic and rodent neural datasets and report improvements in decoding accuracy, particularly under neuronal dropout.

Critical strengths

Framing this as a plug-and-play module makes it sound practically appeal, and the empirical validation across multiple neural datasets and pretrained decoders is a strength. The gains in decoding accuracy are promising in the degraded-observation regime.

Critical weaknesses

There are several major limitations with this paper:

  1. The central methodological contribution is insufficiently justified, both mathematically and in terms of novelty.

The proposed "physical primitives" amount largely to a Gaussian autoregressive prior on motion coordinates, combined with standard manifold operations. It is unclear what genuinely new physical structure is introduced beyond conventional dynamical state estimation.

More importantly, the Bayesian fusion in eq 8 assumes conditional independence between the motion prediction and neural observation. This assumption is not justified: the inferred motion in eq 4 explicitly depends on the same current neural observation. The covariance propagation in eq 7 also treats posterior motion uncertainty as predictive uncertainty without accounting for the relevant dependencies. Consequently, the claimed Bayesian interpretation of the adaptive fusion does not follow from the model as specified.

  1. The experiments do not establish the claimed benefits of the proposed method over appropriate baselines.

The most obvious comparison is missing: a conventional manifold-aware Bayesian filter or smoother with the same motion model and access to the same information. The synthetic torus experiment compares against a Euclidean implementation that mishandles periodic boundaries. Demonstrating that a topology-aware method outperforms one that ignores the topology does not establish a novel contribution.

More seriously, the synthetic experiments supply an analytical velocity term defined using ground-truth motion (Appendix A.1.2 and A.2.2). At zero mismatch, this term is exactly the ground-truth velocity. It is unclear how this information would be available in the intended decoding setting or whether competing methods receive equivalent information. This substantially weakens the experimental conclusions.

Overall, the empirical improvements may reflect the advantages of temporal regularization and correct coordinate handling rather than the proposed physical primitives or variational inference framework. The experiments do not adequately distinguish these possibilities.

Questions for authors

Q1) How is the independence assumption underlying eq 8 justified, given that the motion proposal depends on the current neural observation? How is eq 7 derived under these dependencies?

Q2) How is the analytical velocity input in the synthetic experiments obtained at test time without ground-truth information? Is the same information available to all baselines?

Q3) Why is there no comparison against a standard geometry-aware Bayesian filter using the same dynamics and observations? Without this comparison, what evidence supports the claimed methodological novelty?

3 comments
AuthorsOpenReview ↗

Clarifying the Motion Proposal and Velocity Input, with New Results Against a Geometry-Aware EKF

Thank you for your comments.

Q1/W1: Justification of the Independence Assumption.

Response:

The motion proposal does not depend on the current neural observation, and the conditioning in the original equation was a typo in the time index. In the method and in the released code, the motion proposal is inferred from past observations only, and the current neural observation enters only at the fusion step. Eq 8 therefore fuses a prediction made from the past with the current observation, as in a standard filter, and the motion uncertainty in Eq 7 is predictive uncertainty formed before the observation. The typo is fixed in the revision (Sec. 3.4, Eq. 5), and all reported results are unchanged.

Q2: Source of the Analytical Velocity Input.

Response:

No ground-truth information is used at test time, and the same information is available to all baselines. The analytical velocity term is a fixed, known function of the state, and each method evaluates it at its own estimated state, never at the true state. On the real datasets there is no analytical velocity term at all, so the improvements there do not rely on it.

The appendix wording has been rewritten to make this explicit (App. A.1.2, A.3, A.4), and the simulation code is released: https://anonymous.4open.science/r/MaPIC-F671/

Q3/W1/W2: Comparison with a Geometry-Aware Bayesian Filter.

Response:

This comparison is now included: a geometry-aware EKF on the circle with the same dynamics, the same observations and the same labelled data (App. A.1.4, Table 5), and five filters on rat CA1 (App. A.2, Table 6). At zero mismatch MaPIC and the EKF are statistically indistinguishable; with mismatch and a loss of 80% of neurons, MaPIC reduces error by 39–52%. On CA1, MaPIC is better than the Kalman filter in all 36 paired runs and has the lowest mean error among causal methods on three of four rats, including against KalmanNet.

Beyond conventional dynamical state estimation, our physical primitives are fused with the neural observation through jointly learned uncertainty-dependent weights, and the motion prediction itself is corrected before fusion. A conventional filter has temporal regularization and correct coordinate handling but cannot do this, which is exactly where the gain below appears. This fusion and correction is what lets one module serve diverse settings, across different state-space geometries, multiple real datasets, and different types of frozen base decoders.

MismatchGeometry-aware EKFMaPICError reduction
zero19.38 ± 6.0016.96 ± 5.3512.5%
0.528.62 ± 9.9013.72 ± 2.2652.1% *
1.026.36 ± 9.9714.30 ± 3.7645.8% *
1.522.20 ± 9.5413.56 ± 3.4438.9% *

Median angular error (degrees) during a 2-s loss of 80% of neurons, mean ± s.d. over 10 seeds. * p < 0.05, two-sided paired Wilcoxon test.

Thank you for the detailed response and additional experiments. The new comparisons and ablations address some of my concerns, and I see some practical value in the method, but I don't think the novelty and theoretical foundations are strong enough for ICLR. I am raising my score to weak reject in recognition of the improvements.

Thank you for your recognition. I would still like to explain our core innovations to demonstrate that this work is suitable for ICLR: Our method mainly argues that parsing neural activity into behavior is difficult to describe or specify using linear models, complex dynamics, or data-driven approaches, especially in animals. From a differentiability perspective, any biological behavior can be composed of multiple simple behaviors (especially in terms of movement geometry). Therefore, our method attempts to represent behavior through a linear combination of multiple simple behavioral variations (i.e., primitives) related to biological behavior. This idea differs from the vast majority of previous work and is innovative. On this basis, we derive a variational lower bound to obtain the ideal linear combination. I hope this will help you reassess our method.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Rebuttal

We thank all reviewers for their careful reading, constructive feedback, and continued discussion of MaPIC.

  • Reviewer 7Xrb recognizes the practical value of adding MaPIC to existing decoders, particularly when neural observations are degraded. In their follow-up, they acknowledge the new comparisons and ablations and raise a remaining concern about the method’s novelty and theoretical basis.

  • Reviewer p9Ec sees value in combining temporal correction with the geometry of the decoded variable and highlights the torus boundary analysis as evidence that this design matters near angular boundaries.

  • Reviewer ADta finds the use of physical motion primitives promising and views the gains under neuron loss as evidence of their potential value for robust decoding. After the discussion, they confirmed that our responses adequately addressed their concerns.

To address concern about MaPIC’s novelty and theoretical foundations, we use the torus example to show how MaPIC corrects predicted motion using past observations before a geometry-aware decoder applies its adaptive fusion weights. Suppose the nominal dynamics consistently predict too much rotation along an angular coordinate. Downweighting that prediction increases reliance on the current neural readout, precisely when neuronal dropout may have weakened it. MaPIC can use past observations to correct the predicted rotation before the current readout arrives. It represents local displacement as a linear combination of tangent-space motion primitives and learns a distribution over their coefficients. During training, the trajectory reconstruction term favors displacements that better fit the observed paths, and the KL term regularizes the coefficient distribution toward the dynamical prior. The coefficient means then specify the corrected displacement, and their variances contribute to its fusion weight under the diagonal Gaussian approximation. This distinction motivates both our comparison with a geometry-aware filter given the same information and our separate ablations of the learned correction and the fusion weight.

Additional Experiments and Analyses

  • We compared MaPIC with a geometry-aware circular EKF using the same nominal dynamics, reliability features, labelled data, and validation set.

  • On rat CA1, we compared MaPIC with five filtering and smoothing baselines, including Kalman filters with constant and reliability-based covariance, KalmanNet, a GRU with a Kalman update, and an RTS smoother.

  • We conducted paired, two-sided Wilcoxon tests for the main comparisons.

  • We separately ablated the learned motion correction and the uncertainty-based fusion weight.

  • We released the full code, trained weights, and data in the anonymous repository.

Manuscript Revisions (Main-text changes are highlighted in blue. The revisions and corresponding appendix sections are listed below.)

  • Method description. Section 3 states the variational objective explicitly (Eq. 3, p. 4). Section 3.4 clarifies that MaPIC predicts the motion proposal from history before observing the current readout (Eq. 5, p. 5). Section 3.5 states the fusion approximation, which treats proposal and readout errors as uncorrelated (Eqs. 8–9, p. 5). The algorithm also makes explicit that the GRU memory is updated only after fusion (p. 6).

  • Related work. Section 2 now discusses Cubillos et al. (NeurIPS 2024) and the FALCON benchmark and clarifies MaPIC’s relationship to adaptive-fusion filters (p. 3).

  • Geometry-aware EKF comparison. Section 4 summarizes the comparison (Fig. 4, p. 8), with full results in Appendix A.1.4 and Table 5 (p. 18).

  • Filter comparisons on rat CA1. Appendix A.2 and Table 6 report the results (p. 19).

  • Statistical tests and module ablations. Tables 3–4 report the paired Wilcoxon tests and ablations (pp. 16–17).

Headline Outcomes

Under an imperfect motion model and 80% neuron loss, MaPIC reduces error by 39–52% relative to the information-matched geometry-aware EKF (p < 0.05). With the same geometry and adaptive fusion rule, removing only the learned motion correction increases median angular error from 13.72° to 29.46°. In a separate ablation, replacing the uncertainty-based fusion weight with a fixed weight increases angular error during recovery after signal loss from 4.56° to 5.81°.

On rat CA1, where no analytical motion term is used, MaPIC improves on the Kalman filter in all 36 paired runs and significantly improves the frozen base decoder on all four rats. MaPIC remains effective with only two labelled trials (16 s). On grid-cell data, MaPIC trained on 20% of the data achieves an error of 9.28 cm, compared with 10.16 cm and 9.40 cm for the two baselines trained on the full dataset.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.