acceptodds
Under review as a conference paper at ICLR 2027

PRISM: Population Relational Inference for Streaming Motor Decoding

Abstract

Intracortical brain–computer interfaces must remain robust to session-to-session neural drift while supporting causal, low-latency streaming inference. We introduce PRISM, Population Relational Inference for Streaming Motor Decoding, which separates session interpretation from per-bin prediction. PRISM uses an unlabeled target-session support recording once to construct and cache a Session interface; each subsequent query bin is decoded causally through two Channel SSM blocks, a Population head, HoloSSM, and a motor readout. The decoder maintains fixed-size recurrent state and requires neither target behavior labels nor test-time gradient updates. On the official Few-shot Algorithms for Consistent Neural Decoding (FALCON) private held-out evaluation, PRISM ranks second among few-shot-unsupervised (FSU) entries on all three tasks in the 20 September 2026 leaderboard snapshot, with held-out values of 0.424, 0.574, and 0.272 at official dimensionless normalized latencies of 0.0273–0.0413. On 16 held-out dates of the public Indy reaching dataset, the full model achieves date-averaged . After one-time support calibration, the 262,782-parameter decoder processes each 20-ms query bin in 0.484 ms on one CPU thread. PRISM therefore converts session-level adaptation into a reusable interface while retaining compact, fixed-state, low-latency causal inference.

Primary area:
applications to neuroscience & cognitive science
Keywords:
brain-computer interfaces, neural decoding, cross-session adaptation, state-space models, streaming inference

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_MYV1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PRISM addresses cross-session drift in streaming intracortical motor decoding. It builds a cached, reference-conditioned channel interface from unlabeled support, followed by communicating channel recurrences, population aggregation and complex recurrent memory. Its stated contributions are this computational organization, the relational interface with channel history before final aggregation, and an empirical accuracy–resource characterization. FALCON results, 16-date Indy experiments and resource measurements support a useful complete system; the incremental benefit of relational processing and the preferred placement of recurrence are less clearly established.

Critical strengths

S1. Useful performance at modest streaming cost. PRISM achieves competitive cross-session decoding performance on FALCON with lower reported latency than the displayed few-shot unsupervised comparators. Its low computational cost and compact recurrent state support its practical value for streaming motor decoding.

S2. Useful controlled observations. Controlled experiments demonstrate the benefits of target-session calibration and channel recurrence within PRISM.

Critical weaknesses

W1. The motivation for the specific design and its incremental contribution are insufficiently distinguished from prior work. The cross-session decoding problem and the contributions are listed, but it remains unclear why the proposed interface and recurrence organization are preferable to existing unsupervised adaptation and temporal/population modeling approaches. SPINT [1] already derives unit identities from unlabeled support and processes causal per-unit histories; EIT [2] models individual temporal dynamics before population interaction. These connections warrant a clearer account of what its particular interface and organization add. The full-model comparison with SPINT does not isolate an advantage over IDEncoder.

The ablations leave this distinction unresolved. NoTargetCalibration retains the learned interface and 108-second replay while substituting source-fixed normalization/support quantities , so its performance drop does not isolate relational matching. Table 6's existing StandardizationOnly comparison uses one checkpoint per model and incompletely specifies the bypass. Although the 51-feature construction is described, the existing controls do not isolate the incremental contribution of its relational feature groups. A clearly defined, replicated StandardizationOnly comparison is the smallest useful check; more targeted feature-group or IDEncoder-style controls would be informative.

POSSM [3] is also a relevant, cited recurrent decoder, but is absent from the measured comparisons. The current ablation removes channel-level recurrent memory rather than relocating it, so it does not isolate the effect of recurrence placement. A POSSM-style or another matched aggregate-first recurrent baseline would address that distinction. The existing evidence therefore supports the usefulness of channel recurrence within PRISM, but does not establish an advantage of the proposed placement.

W2. The methodology and supporting figures do not explain the contribution clearly. Many equations and variables are introduced without a sufficiently clear account connecting the support features, learned encoder, channel map and recurrent blocks to source training and target-session inference. Figure 2 does not adequately explain these operations or which quantities are learned, computed from support, cached or updated per bin. A revised diagram and a short training/calibration/inference walkthrough would make the design and its claimed contribution easier to assess.

The supplementary presentation is also unfinished. S1 includes the instruction, “The caption should identify the displayed recording or date, the query interval, the velocity components, and the meanings of all lines, markers, and shaded regions.” S2/S4 contain similar editorial directions; S3's plot and caption describe different tasks/context settings. Together with the unclear methodological exposition, they weaken confidence in assessing the mechanism claims; caption cleanup alone would not resolve the main scientific concern.

References

[1] Trung Le et al. SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding. NeurIPS, 2025.

[2] Ran Liu et al. Seeing the forest and the tree: Building representations of both individual and collective dynamics with transformers. NeurIPS, 2022.

[3] Avery Hee-Woon Ryoo et al. Generalizable, real-time neural decoding with hybrid state-space models. NeurIPS, 2025.

Questions for authors

Q1. What motivated the selection of the 51 support features, and what evidence supports the contribution of the relational feature groups? What distinctive capability does the shared encoder/reference interface offer compared with SPINT’s IDEncoder, and what evidence supports it? Please clarify whether relational feature engineering supplies the claimed distinctive benefit.

Q2. Beyond the Full-versus-NoTargetCalibration bundle comparison, what evidence distinguishes the contribution of the learned interface from target-session normalization? In particular, what exactly does StandardizationOnly disable? Table 6 shows nearly identical clean performance but positive mean gains for Full under persistent channel loss. How do the authors interpret this difference, and does the robustness benefit persist across training seeds?

Q3. Can the placement claim be supported by a POSSM-style or another comparable-capacity aggregate-first recurrent baseline using the same interface, supervision and calibration access, with resource differences reported? The current ablation removes channel-level recurrent memory rather than relocating it, so it does not isolate recurrence placement. Alternatively, is the intended claim only that channel recurrence helps this architecture?

Q4. What explains PRISM's latency advantage, and how does it compare with one suitable streaming decoder, such as POSSM or a Mamba-based decoder, with accuracy, capacity and context differences reported? Please identify the exact implementation used for timing and clarify what evidence attributes the advantage to the proposed organization rather than implementation or model-size differences.

Q5. The supplied local trainer differs from Appendices A.1, A.4 and A.5 in source-reference selection, the training objective, learning-rate schedule and backpropagation procedure, prefix-state handling during training, and checkpoint selection. For example, train.py selects the first training session as the reference rather than applying Eq. 16, and saves the final model without evaluating the allocated validation sessions. Please identify which implementation and configurations generated the reported results and explain their correspondence to the supplied package.

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

This paper introduces PRISM, a streaming motor decoding method leveraging unlabeled neural data from target sessions for neural adaptation. The method is ranked two on three Falcon benchmarks, demonstrating improved performance on another public dataset over baselines.

Critical strengths
  • Demonstrate how to utilize unlabeled data from target session for adaptation
  • Evaluated and ranked second on public Falcon benchmark on few-shot unsupervised learning
  • Outperformed baselines on another public monkey dataset, with ablation studies
Critical weaknesses
  • Missing comparison with rank one methods on Falcon benchmarks
  • Proposed method is complicated with many hyperparameters, a more comprehensive ablation study is likely needed
  • In the introduction section, I find the motivation of “similar behavior and different neural activity” unclear and rather abrupt. I don’t see how describing Falcon benchmark and Indy dataset helps motivating the story.
Questions for authors
  1. Why is the number one method from falcon benchmark (ABACUS) skipped from the comparison? And why is SPINT not evaluated in Indy datasets? Also in general what are the strategies for FSU among the three methods? (ABACUS, SPINT and PRISM)
  2. I am surprised there are nearly no cross-day variance in Indy datasets, in fact, baseline like NoMADs are very variable and it can be better than the proposed methods in some days (judging by the std). Is there just a few day outlier? Can the author provide statistic test for results in Indy?
  3. There are many equations included in the manuscript, which parts are the novelty of this work? (e.g. HoloSSM?)
  4. How are reference channels selected? What is the size?
  5. In general I think the paper would benefit from a more thorough ablation study, like number of banks in holoSSM, number of SSM layers, and reference size. Are those value hyperparameters tuned?
  6. In equation (3) are keys and values the same vectors? Or are there separate projection
  7. I believe it could be beneficial to talk about the figure 1(b) in the introduction.
Reviewer_S2pfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

PRISM is a motor decoding approach that separates session interpretation from streaming prediction. A per channel and window descriptor of 51 dimensions is constructed reflected local activity statistics and lagged associations. This is fed to a tokenizer. A similarity matrix between target and reference channels is extracted.

A state-space model for long-sequence recurrence gives each channel its own temporal state. A population head then is converting channel features into one vector for motor decosing. This is fed to a multiscale complex-valued memory (HoloSSM) which is predicting the signal to be supervised. The deployment has a support phase and a query phase.

Critical strengths
  • The distinction between one-time session interpretation and per-bin causal decoding directly addresses neural drift and real-time deployment constraints

  • The calibration ablation shows degradation when removing target calibration, same for removing channel recurrence or population recurrence.

  • Extensive parameter reporting

Critical weaknesses
  • The paper is really hard to read. A plethora of choices are introduced without any motivation making the paper look like a brute-force engineering solution.

  • While some ablations give insight (target calibration) all other choices like normalization, soft channel mapping amd gain/bias corrections are not explained either intuitively or experimentally.

  • It is not clear what is the novelty: Right now the way the paper is written sounds like a bunch of engineering without following any decoding or information-theoretic principles.

Questions for authors
  • Add a staged ablation separating: target normalization only, normalization + mapping, mapping without gain/bias, gain/bias without mapping, full interface

  • Provide full explanation for introducing them. A hypothesis that these components improved when stated should be based on some intuition or previous findings.

  • State the novelty in one precise paragraph, contrasting PRISM with NoMAD, SPINT, and causal SSM decoders.

Comments on the paper

2 comments

Dear Reviewers,

Please take a moment to review the authors’ rebuttal, assess whether your concerns have been adequately addressed, and actively participate in the discussion. Please update your assessment if needed. Thank you for your time and effort!

AuthorsOpenReview ↗

Closing Summary

We thank the reviewers and Area Chair for the discussion. The original reviews recognized PRISM's competitive decoding performance and low streaming cost, while raising questions about design motivation, component contributions, and implementation clarity. We summarize the additional evidence and completed manuscript revisions below.

1. The design and implementation are now explicit.

We added the requested staged interface controls separating target normalization, residual mapping, and gain/bias correction, and explained the session changes that motivate these operations. We clarified the relationship to NoMAD, SPINT, EIT, and causal SSM decoders. Figure 2 itself has been redrawn to distinguish support-time computation, cached quantities, and per-bin state updates. Section 3 now provides a training–calibration–inference walkthrough, and Table 3 identifies saved-snapshot entry points and run-record fields for the reported procedures.

The earlier descriptor and aggregation-order controls have also been restored in Appendix B.10, Tables 11–12. Their actual configurations are stated explicitly, with the historical AggregateFirst runs distinguished from the corrected full-interface follow-up.

2. The discussion strengthened the experimental evidence.

The three-seed calibration analysis shows improvements of 0.2980 in All and 0.5474 in Far R2R^2, with positive paired differences on 14/16 test dates. Beyond target normalization, the complete interface improves All R2R^2 under persistent channel loss by 0.0153, 0.0270, and 0.0191 at 25%, 50%, and 75% removal, winning on 14/16, 14/16, and 16/16 paired dates. The revised presentation distinguishes training-seed variability, recording-date variation, and mask variability.

In the follow-up discussion, we also reported single-seed controls for SSM depth, bank count, reference size, and SPINT-style support conditioning within the PRISM decoder. The depth control shows a 0.0506 All R2R^2 gain from the second channel SSM layer, with improvements on 15/16 dates. These experiments provide evidence about the architecture settings and support-processing choices.

3. Contribution supported by the revised submission.

PRISM turns unlabeled session support into a cached, reference-conditioned interface that subsequent neural bins reuse during fixed-state causal decoding. This organization addresses adapting to a changed recording while preserving a compact streaming prediction path.

The method ranks second among FSU entries on all three FALCON tasks in the reported 20 September 2026 leaderboard snapshot. On 16 Indy test dates, it achieves mean All R2=0.5514R^2 = 0.5514 across three training seeds. After calibration, the reported 262,782-parameter decoder processes each 20-ms bin in 0.484 ms on one CPU thread, with 63 KiB of recurrent state.

Our contribution is supported by the complete decoder's official benchmark performance, controlled calibration and channel-loss results, and measured streaming costs. We believe this combination makes a substantive contribution to cross-session neural decoding. We thank the reviewers and Area Chair for the discussion that helped us strengthen both the evidence and its presentation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.