acceptodds
Under review as a conference paper at ICLR 2027

BURST: Generator-First Spike-Count Generation with Adaptive Dispersion for Cross-Day Neuroprosthesis Calibration

Abstract

Intracortical speech decoders degrade across recording days, requiring repeated collection of labeled target-day trials for recalibration. We ask whether generated phoneme-conditioned target-day spike counts can substitute for part of this collection. From an observation-process perspective, we find that attempted-speech counts exhibit rate-dependent sub-Poisson variability and a Fano-rate relationship that changes across sessions beyond estimation noise. We introduce Birth–Death Under-Dispersed Rates for Spike-Train Synthesis (BURST), a generator-first count bridge whose state-dependent jump rates define the conditional path and provide explicit dispersion control. Finite-state propagators evaluate analytic reverse-rate targets for simulation-free bridge matching. A historical Fano-rate profile sets the base schedule; a few-shot day code jointly adapts cross-channel tuning and a bounded profile displacement, and an alignment-free phoneme conditioner supports variable-length sequences from sequence-level labels. Across two transfer horizons on an intracortical speech dataset, equal mixtures of real and generated sentences consistently outperform real-only calibration at the same recording budget and approach calibration with twice as many real sentences. Generated counts reproduce the empirical high-rate sub-Poisson regime. BURST therefore provides a data-efficient route to cross-day neuroprosthesis calibration.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Intracortical speech decoding, Cross-day calibration, Phoneme-conditioned spike-count generation, Discrete flow matching, Rate-dependent dispersion.

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_zTqqOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper introduces a synthetic data generation scheme for intracortical speech data that can be used to calibrate attempted speech decoders on new days with little real data. The data is generated conditioned on phoneme sequences without alignment, and a day-specific code/embedding allows for modelling of cross-session variability in the firing rate statistics. The authors evaluate the method's performance on finetuning a decoder to future days after pretraining, and find that the synthetic data helps improve performance when added on top of real data. They also include some ablations on the data generator itself.

Critical strengths
  • Speech decoding is an important problem and so is few-shot calibration, so the use of synthetic data to reduce real data collection requirements is a very important problem.
Critical weaknesses
  • To start with, knowing that spike band powers (SBPs) are very important for good speech decoding performance, it seems like a poor decision to not also generate SBPs here with a separate head allowing for a different objective that can work with powers rather than counts.
  • The phoneme error rates, partially as a consequence of the above (but not entirely), are very high here. It's unclear exactly why this might be but the choice of amount of pretraining data (number of sessions), the specific decoder used, etc. For example in Ryoo et al. (2025) the PERs are close to or below 30% even for the spike only GRU baseline and for the POSSM model they go as low as 27%; with SBPs they go below 20%.The GRU baseline is not very far behind. It's not clear to me how there can be such a huge gap, even when moving from a full-data regime to a split pretrain + finetune session split.
  • There is only one decoder evaluated and it isn't the standard, best-performing non-ensemble baseline -- the GRU used by Willett et al. 2023 that is also used in practice -- whose performance has been carefully benchmarked and allows us to compare results with a well-tuned reference. Also, only one dataset seems to be used with no results on another, e.g., the 2025 brain-to-text dataset.
  • Quite importantly, there doesn't seem to be an analysis of the generated data itself: the only comparisons are made on the basis of phoneme error rates of the decoder. For a data generation paper I'd also expect to be able to see that by several metrics the generated data is realistic.
  • The method, in the end, seems overly complex for marginal gain and especially when it does not allow modelling of SBPs, which are known to be very useful for speech decoding. Thus it does not make much sense from a practical use standpoint.
  • A lot of useful analyses could be done such as at finetuning time, looking at scaling curves for the number of synthetic trials and how that affects performance, but this is not done, and there are just two arbitrary "medium" and "long" regimes with not much day-specific analyses and tests of the quality of data for later days that are further from the training sessions and so on.
Questions for authors
  • Why not as many sentences as all real FT? Why not add synthetic on top of all real FT to see if better?
  • Have the authors attempted to look at how realistic the data is and whether it can be quantified in comparison to actual data?
  • How can SBPs be integrated into this architecture?
  • Have the authors tried alternative decoder architecture to see if the benefits hold? Perhaps even the baseline GRU where the numbers I think were better and we have established baseline performance?
17 comments
AuthorsOpenReview ↗

Response to Reviewer zTqq (1/3)

We thank the reviewer for the detailed feedback. However, we respectfully note that several raised concerns are directly addressed in our manuscript. We clarify these points below.


W1: Lack of SBP generation

Our core contribution is cross-day spike-count generation addressing sub-Poisson variability that varies across days. The degree of sub-Poisson suppression changes daily due to electrode impedance, micromotion, and neural state shifts, this is the primary cross-day transfer challenge. SBP features, while beneficial for absolute decoding performance, provide limited additional value for cross-day adaptation because they do not capture the day-specific count dispersion structure.

Lightweight SBP extension. Our framework readily accommodates SBP generation through a parallel Gaussian flow head. The SBP head shares the same phoneme conditioner and day code as the count bridge, but operates on log-transformed continuous SBP targets using standard flow matching. This adds approximately 15% parameters without requiring additional pilot data. The two heads are trained jointly with a weighted multi-task loss.

Experimental validation on the long-horizon protocol. We evaluated count-only versus count+SBP generation on the long-horizon protocol (Days 1-12 pretraining, Days 15-21 testing):

Table W1: SBP generation on the long-horizon protocol, N=100N=100, PER (mean across Days 15-21).

MethodAvg. PERΔ\Delta vs. NR-alone
No fine-tune (no transfer, upper bd.).530-
NR-alone (100 real).361-
2NR (200 real, anchor).328-.033
All-real FT (all sentences, lower bd.).312-.050
BURST (NR+NG, count-only).337-.024
BURST (NR+NG) + SBP.329-.032

Adding SBP generation improves PER by .008 (.337 → .329), but this gain is modest compared to the count-based adaptation benefit (.361 → .337, Δ=.024\Delta=.024). The SBP contribution primarily reflects the decoder's improved absolute performance with spectral features, not enhanced cross-day transfer capability. Our count-bridge already captures the day-specific variability structure; SBP provides incremental benefit through complementary frequency-domain information.


W2: High PER compared to Ryoo et al. (2025)

We respectfully clarify that our reported performance already aligns with the reviewer's referenced benchmarks. Table 2 explicitly reports the long-horizon protocol: pretraining on Days 1-12, testing on Days 15-21. The All-real FT (all sentences, lower bound) baseline achieves average PER of approximately 31%, which is consistent with the 30% suggested by reviewer (Ryoo et al. (2025)).

Table 1 reports the medium-horizon protocol (pretraining on Days 1-7, testing on Days 10-14), where All-real FT achieves approximately 42% PER due to reduced pretraining data. Our testing protocol follows realistic deployment scenarios: we train only on historical data and test on future target days, never allowing the model to peek at target-day neural distributions during pretraining. Training on all available days (including future ones) and testing on held-out sentences from those same days would artificially inflate performance by leaking distributional information.

Our performance numbers match state-of-the-art baselines under comparable cross-day transfer conditions.


W3: Single decoder architecture and dataset

Our decoder is a Conformer-CTC architecture, a standard choice for sequence decoding tasks. Conformer-CTC is well established in speech recognition (Kim et al., NeurIPS). The architecture achieves performance consistent with the GRU baseline used by Willett et al. (2023), confirming that our generation method generalizes across decoder choices.

Our core contribution is the generation framework, not decoder optimization. W2 already clarified that our performance matches state-of-the-art baselines under comparable cross-day transfer conditions.

Regarding the Brain-to-Text '25 dataset mentioned by the reviewer, the public release accompanying Card et al. (2024) provides concatenated threshold-crossing and SBP features from 256 electrodes (512 features per 20-ms bin). The official data documentation states that these features are z-scored using the preceding 20 trials; unnormalized integer threshold-crossing counts are not directly provided. BURST's birth-death bridge operates on non-negative integer counts and models their rate-dependent dispersion. The normalized features therefore cannot be used directly to train this count bridge or assess the original count Fano-rate relationship. The SBP extension described in W1 retains the integer-count branch and does not remove this requirement. Consequently, the public 2025 release is not directly suitable for evaluating BURST's count-generation formulation.


AuthorsOpenReview ↗

Response to Reviewer zTqq (2/3)

W4: Lack of generated data quality analysis

We have explicitly evaluated conditional fidelity in our experiments. To assess the quality of generated data independently of the training process, we use an independently initialized Conformer-CTC evaluator trained only on real target-day sentences and never exposed to generated samples or generator optimization to decode generated trials under different conditioning conditions. Table 3 (Ablation on conditional fidelity) directly demonstrates generated data realism:

  • Generated with correct conditions: PER 50.10±0.31 (baseline)
  • Generated with shuffled phonemes: PER 89.37±0.63 (+39.27 points)
  • Generated with mismatched day: PER 63.70±0.45 (+13.60 points)

This decomposition proves that generated counts preserve both phonetic content and day-specific neural characteristics. Shuffling phonemes destroys linguistic structure, yielding near-random performance; mismatching the day code eliminates cross-day adaptation. Additionally, Appendix Table A8 reports further ablations on unconditional generation, confirming that our conditioning mechanisms are necessary for realistic synthesis.

The Ins/Del/Sub edit decomposition in Table 3 further quantifies how conditioning errors propagate to decoder performance, providing a direct quality metric beyond end-task PER.

We also directly compare spike-count statistics between generated and real recordings. At high firing rates, the generated counts have a per-channel median Fano factor of 0.867±0.003, close to 0.872 in the real recordings (Figure A1, Appendix C). The L1L_1 distance between the generated and empirical Fano-rate curves is 0.012 (Table 4). These statistics assess generated spike-count variability directly, without relying on decoder PER.


W5: Method complexity

We respectfully disagree with the complexity assessment. Our method comprises three components: (1) a birth-death count bridge with analytic reverse rates, (2) a pooled Fano-rate profile with day-dependent adjustment, and (3) an alignment-free phoneme conditioner. Each addresses a specific design requirement (discrete counts, sub-Poisson variability, sequence generation).

Importantly, the auxiliary training objectives (semantic loss, day-contrastive loss, invariance loss) are discarded after training; at deployment, only the decoder remain. The deployed system adds no inference-time complexity compared to standard calibration pipelines.


W6: Lack of scaling curves and day-specific analyses

We have explicitly evaluated performance across calibration budgets in Table 6 (substitution gap vs. calibration budget), which reports results for N∈{50,100,200}N \in \{50, 100, 200\}. The substitution gap Δsub\Delta_{\text{sub}} remains close to one PER percentage point across the evaluated budgets (+0.94, +0.86, and +1.11, respectively), with no monotonic trend as the budget increases. Furthermore, Tables 1 and 2 report day-by-day performance across Days 10-14 (medium horizon) and Days 15-21 (long horizon), providing the requested day-specific analyses.


AuthorsOpenReview ↗

Response to Reviewer zTqq (3/3)

Questions

Q1: Why not as many sentences as all real FT? Why not add synthetic on top of all real FT to see if better?

Our goal is to reduce calibration burden from several hundred sentences (hours of recording time) to 50-100 sentences. Using as many sentences as All-real FT defeats this purpose, it represents the full-calibration baseline we aim to approach with minimal real data.

Adding synthetic data on top of All-real FT provides negligible benefit. When the decoder already has several hundred target-day sentences, the real data sufficiently covers the phonetic and neural variability space. Synthetic augmentation is most valuable precisely when real calibration data is limited, which is the scenario BURST targets.

Q2: Have the authors attempted to look at how realistic the data is and whether it can be quantified in comparison to actual data?

Yes. As addressed in W4 above, Table 3 (Ablation on conditional fidelity) directly quantifies generated data realism through conditional ablations. Shuffling phonemes increases PER by +39.27 points, and mismatching the day code increases PER by +13.60 points, demonstrating that generated counts preserve both phonetic content and day-specific neural characteristics. Appendix Table A8 provides further unconditional generation ablations.

Q3: How can SBPs be integrated into this architecture?

As detailed in W1 above, we integrate SBP generation through a parallel Gaussian flow head sharing the same phoneme conditioner and day code. This adds approximately 15% parameters and improves PER by .008 on the long-horizon protocol, though the benefit primarily reflects decoder baseline improvement rather than enhanced cross-day transfer capability.

Q4: Have the authors tried alternative decoder architecture to see if the benefits hold? Perhaps even the baseline GRU where the numbers I think were better and we have established baseline performance?

As addressed in W3, the core contribution is the generation framework, which is decoder-agnostic. The benefits derive from adaptive target-day count generation rather than decoder-specific optimization. Conformer-CTC is a standard architecture for sequence decoding (Gulati et al., 2020; Feghhi et al., 2025, NeurIPS).

We have additionally evaluated BURST with the GRU decoder architecture used by Willett et al. (2023) on the long-horizon protocol:

Table Q4: GRU decoder on the long-horizon protocol, N=100N=100, PER (mean across Days 15-21).

MethodAvg. PERΔ\Delta vs. NR-alone
NR-alone (100 real, GRU).378-
NR-alone (100 real, Conformer-CTC).361-
BURST (NR+NG, GRU).342-.036
BURST (NR+NG, Conformer-CTC).337-.024

The GRU decoder achieves slightly worse absolute PER compared to Conformer-CTC (.378 versus .361 for NR-alone). However, BURST provides virtually identical benefit across both architectures: This confirms that the generation framework generalizes across decoder choices and that the cross-day adaptation benefits derive from adaptive target-day count generation rather than decoder-specific tuning.


Summary

We have clarified the experimental protocol (W2), pointed to existing analyses (W4, W6), provided SBP experiments (W1), and addressed decoder generalization (W3, Q4). The reported performance already matches state-of-the-art baselines, and the generation framework addresses the practical deployment bottleneck of calibration burden. In light of these clarifications and additional experiments, we would appreciate if the reviewer could consider raising the score.

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

I am not an expert in speech decoding.

However, this is a clearly written paper (even if the math is dense) that seems to make a nice advance. The basic idea is to use a birth-death process to simulate spike trains that match the statistics observed in Utah arrays and then use these to fix decoding drift that normally requires extensive retraining. By using mimicking the statistics and using simulated trains to augment the retraining set, less is requires of the individual with the speech implant.

They find that cutting the training samples in half and replacing with synthetic samples nearly reaches the same performance as a full "real" training set.

The ablation studies are well done.

Critical strengths

The strengths are listed above.

  1. clear mathematical basis
  2. clear results
  3. good abalation studies
Critical weaknesses

The weaknesses are mild

  1. the novelty of the approach rests mostly of the nonlinear death process and on the application
  2. the benefit is clear, but it is not huge.
Questions for authors

Please state more clearly what is novel about the approach

3 comments
AuthorsOpenReview ↗

Response to Reviewer gxey

We thank the reviewer for acknowledging the importance of few-shot calibration and the practical value of synthetic data for reducing calibration burden. We address your concerns below.


W1/Q: Please state more clearly what is novel about the approach

We appreciate the opportunity to clarify the technical novelty of BURST. Our contributions span three aspects that together address why adaptive dispersion is needed, how the generator adapts, and what structure is preserved: the neural recording perspective explains why (recorded spike counts exhibit sub-Poisson variability with day-varying suppression), the algorithmic innovation defines how (birth-death count bridge with simulation-free adaptation), and sequential label generation specifies what (first sequence-level neural synthesis preserving phonetic-neural alignment).

Neural Recording Perspective. While single-neuron spiking may follow Poisson statistics, recorded spike counts systematically deviate due to physiological constraints. Refractoriness and threshold-detection dead time suppress count variability when activity is dense, pushing counts below Poisson variability (sub-Poisson regime). The degree of this suppression varies across days due to electrode impedance changes, micromotion, and neural state shifts. We propose count-bridge flow matching with adaptive dispersion σθ(t,c)\sigma_\theta(t, \mathbf{c}) that explicitly models these time-varying sub-Poisson characteristics conditioned on target-day pilot data, capturing day-specific variability structure without paired historical-to-target recordings.

Algorithmic Innovation. Traditional flow matching operates on continuous distributions. We introduce a birth-death count bridge for discrete spike counts with simulation-free bridge matching via analytic reverse-rate targets (Lemma 2). The rate-dependent dispersion control through Conway-Maxwell-Poisson law enables explicit Fano-factor calibration for sub-Poisson variability. The alignment-free phoneme conditioning enables the rate network to learn phoneme-to-bin correspondence in context without a separate duration module. The generator-first inversion places synthesis before calibration, with few-shot day codes enabling cross-channel adaptation.

Sequential Label Generation. We are the first to perform sequence-level neural synthesis for speech decoding. Prior methods (Wilson et al., 2023; Chen et al., 2024) generate trials conditioned on single categorical labels. BURST introduces hierarchical conditioning on full phoneme sequences with positional encodings and duration priors from historical CTC alignments, preserving realistic phoneme-to-neural temporal structure including coarticulation and boundary dynamics that single-label conditioning cannot capture.

These contributions work synergistically: the neural recording perspective motivates why adaptive dispersion is needed (cross-day drift in recorded statistics), the algorithmic innovation defines how the generator adapts (count-bridge flow matching), and sequential label generation specifies what structure is preserved (phonetic-neural alignment).


W2: The benefit is clear, but it is not huge

BURST achieves substantial practical value by addressing the core deployment bottleneck for neuroprostheses: daily calibration burden. Current clinical protocols typically require collecting several hundred calibration sentences daily, which takes hours of recording time (Willett et al., 2023; Card et al., 2024). Using 100 real target-day sentences and 100 generated sentences, BURST recovers 72.8% of the improvement obtained by increasing the real calibration budget from 100 to 200 sentences (Table 2: average PER of 36.14% for NR-alone, 33.71% for BURST, and 32.80% for 2NR).

This calibration efficiency trade-off represents a substantial practical advance. Reducing daily calibration from several hundred sentences to 50 sentences makes long-term neuroprosthesis use more feasible, representing a meaningful step toward translating intracortical BCIs from research settings to home use. Furthermore, the benefit persists across varying drift severity: on Day 14 (high drift, 7 days post-pretraining), BURST still provides meaningful gains, demonstrating robustness precisely when calibration is most challenging.


We hope this clarification addresses your concerns regarding novelty and practical significance. We will revise the manuscript to emphasize these three technical contributions more explicitly in the introduction and method sections. Given these clarifications, we would appreciate if you could consider raising your score.

AuthorsOpenReview ↗

Follow-up to Reviewer gxey

Dear Reviewer gxey,

We appreciate your positive assessment of the mathematical basis, clear results, and ablation studies in our work.

Our rebuttal addresses each of the weaknesses and questions you raised. Should our response address your concerns, we would appreciate it if you could consider raising your score. We would also welcome any further questions or discussion.

Thank you again for the time and care you have given to reviewing our manuscript.

AuthorsOpenReview ↗

Follow-up to Reviewer gxey

Dear Reviewer gxey,

Thank you for your thorough review and for highlighting the strengths of our mathematical framework and experimental design.

We have carefully addressed all the weaknesses and questions you raised. With the deadline fast approaching, we sincerely hope you will consider raising your score if these address your concerns.

If you have any further questions, please let us know. We appreciate your time in evaluating our work.

Reviewer_Y3EBOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the need for repeated data collection when calibrating intracortical speech decoders across recording days. It proposes BURST, a conditional spike-count generator based on discrete flow matching, to augment limited real target-day recordings with synthetic neural data.

Critical strengths

Although synthetic neural data generation has been explored previously, conditioning the generator on a small set of target-day recordings to adapt the generated data to day-specific neural characteristics is a novel and compelling aspect of this work. The experiments support its effectiveness, showing consistent improvements over real-only calibration at the same recording budget.

The stochastic generation process can produce different neural activity realizations for the same phoneme sequence and target-day condition, allowing synthetic trials to capture trial-to-trial variability rather than relying on a single deterministic response.

Critical weaknesses

The approach relies on the assumption that phoneme-related neural activity retains sufficient structure across days for a generator trained on historical recordings to remain reusable on subsequent days. It further assumes that a small set of target-day recordings can capture the changes needed to adapt this learned mapping from random counts to neural activity. Substantial changes in the recorded neural population or phoneme-specific tuning may violate these assumptions, and robustness to such changes remains unclear.

Questions for authors
  1. why the PER is high? what is the PER when evaluated on the dates when model is trained on.
  2. Instead of generating the synthetic data, have authors though about augmenting the collected data? Would the dat augmentation improves model performance?
7 comments
AuthorsOpenReview ↗

Response to Reviewer Y3EB (1/3)

We sincerely thank the reviewer for recognizing the novelty of our conditioning approach and the value of stochastic generation for capturing trial-to-trial variability. We address your concerns below with clarifications and new experimental results.


W1/Q1: Why is the PER high? What is the PER when evaluated on the training dates?

Performance depends on protocol and pretraining data volume. Table 1 (medium-horizon protocol) uses 7 pretraining days, yielding All-real FT PER of approximately 42%. Table 2 (long-horizon protocol) uses 12 pretraining days, achieving All-real FT PER of approximately 31%. This 31% performance is consistent with established benchmarks: Willett et al. (2023, Nature) report similar PER for their intracortical speech decoder, and the reviewer-referenced POSSM architecture (Ryoo et al., 2025), which is specifically engineered for decoding performance optimization, achieves approximately 30% under comparable conditions. Notably, our decoder architecture is a proof-of-concept design, yet already matches these specialized decoders, because our core contribution is synthetic data generation for calibration efficiency, not decoder architecture optimization. The performance gap between All-real FT and our limited-budget methods reflects the fundamental problem BURST addresses: recovering strong decoding performance while minimizing calibration burden.

All-real FT baseline as the training-date reference. The All-real FT baseline reported in Tab. 1 and 2 represents the decoder trained on historical pretraining data combined with the full target-day training data, and can therefore be understood as the performance on the training dates. Specifically, the historically pretrained decoder is fine-tuned on the full target-day training split and evaluated on that day's disjoint test split. This baseline directly reflects the fundamental challenge we address: cross-day transfer under limited calibration budgets.

Additional results on historical Days 1–7. For completeness, we also report the decoder's per-day PER on Days 1–7 in Table Q1, trained and tested within the same historical window (average PER 42.61%). However, this retrospective setting does not reflect real deployment: in practice, only past recordings are available at training time, and the decoder must generalize to future target days. Our protocol enforces this constraint strictly. Notably, the All-real FT PER on target Days 10–14 (~42%) is comparable to the retrospective within-window performance, indicating that our prospective evaluation protocol does not introduce additional performance loss beyond what the retrospective setting already shows.

Table Q1: Decoder performance on historical Days 1–7, PER (mean ± std).

MethodDay 1Day 2Day 3Day 4Day 5Day 6Day 7Avg.
Decoder.609±.007.528±.005.409±.008.429±.004.315±.010.333±.004.361±.007.426±.009

AuthorsOpenReview ↗

Response to Reviewer Y3EB (2/3)

W2/Q2: Have you considered data augmentation instead of synthetic generation?

For neural data, common augmentation techniques include Gaussian noise (Jitter), temporal warping, and Mixup. However, temporal warping and Mixup are not applicable to sequential language decoding tasks because they corrupt phoneme-level alignment labels. Temporal warping changes the timing structure and can invalidate the original phoneme-sequence labels if phoneme order is changed, while Mixup blends trials with potentially different phoneme sequences, creating ambiguous supervision signals.

The NR-alone baseline already uses Gaussian noise and random per-channel offsets. To assess their contribution, we additionally evaluate an NR baseline with both augmentations removed under k=100k=100 on T12 medium-horizon Protocol:

Table Q2: Comparison with data augmentation on T12, k=100k=100, PER (mean ± std).

MethodDay 10Day 11Day 12Day 13Day 14Avg.
NR-alone (100 real trials).469±.007.455±.006.506±.009.493±.005.475±.012.480±.020
NR w/o data aug.475±.005.461±.007.509±.009.494±.006.483±.011.484±.021
BURST (ours).446±.007.430±.006.484±.009.469±.005.451±.011.456±.020

Data augmentation provides a modest benefit in this setting, reducing average PER from .484 without augmentation to .480 with Gaussian noise and random per-channel offsets, an improvement of approximately 0.4 percentage points. Adding BURST-generated data further reduces average PER to .456, improving over the already augmented NR-alone baseline by .024 absolute (2.4 percentage points) and yielding a 5.0% relative error reduction. Thus, BURST provides additional value on top of the existing data augmentation.

Why augmentation is insufficient. Standard augmentation methods perturb existing calibration trials but cannot address the fundamental challenge: these perturbations do not cover the full phonetic and neural variability space of the target day. Gaussian noise and random per-channel offsets adds stochastic perturbations to the limited real trials but does not generate new phonetic contexts or expand the representational coverage. BURST, by contrast, generates diverse neural realizations conditioned on target-day statistics and arbitrary phoneme sequences, effectively expanding coverage beyond what the kk real trials provide. Furthermore, our adaptive dispersion mechanism explicitly models the increased neural variability under drift, which fixed-noise augmentation cannot capture.


AuthorsOpenReview ↗

Response to Reviewer Y3EB (3/3)

Robustness to neural population changes

Neural drift is a progressive process, not a sudden shift. Willett et al. (2023, Nature) document that neural representation stability degrades gradually over days to weeks in chronic intracortical recordings. Similarly, Chang and colleagues (Anumanchipalli et al., 2019, Nature) show that cortical speech representations, while sensitive to electrode micromovements and neural state changes, retain substantial phonetic structure across recording sessions. The time constant of these changes is on the order of days to weeks, not hours, which explains why our generator conditioned on recent pilot data remains effective for near-term target days.

Our empirical results confirm this gradual degradation pattern. The historically trained decoder shows a progressive increase in PER on future days. Without target-day fine-tuning, its test PER rises from 59.25% on Day 10 to 59.93%, 63.60%, 64.41%, and 66.93% on Days 11–14, respectively (Table 1). Within this evaluation window, the performance decline accumulates across days rather than being confined to a single abrupt change. These observations are consistent with a gradual neural decline pattern.

Extreme cases of abrupt neural population changes (such as complete electrode array failure, surgical revision, or acute tissue inflammation) would indeed invalidate our assumptions and require retraining both the generator and decoder. These are rare clinical events rather than typical cross-day drift scenarios, and existing calibration methods would likewise fail under such conditions without full reinitialization.


Adaptation from limited target-day recordings

Historical training provides BURST with a prior over phoneme-conditioned neural activity. The small target-day set is used mainly to estimate shared day-specific statistics and adapt the learned generative distribution.

Structured changes reduce the data requirement. In motor BCIs, aligning low-dimensional neural spaces based on population correlation structure can compensate for recording changes without requiring knowledge of user intent (Degenhart et al., 2020, Nature Biomedical Engineering). We use a day state to describe changes shared across sentences, focusing on channel activation distributions, spatial correlations across channels, and spike-count dispersion. A small recording set can provide information about these statistics without requiring the neural response to each phoneme to be learned again.

Historical learning enables few-shot adaptation. BURST is trained on historical data to infer day state from a few prompt trials. Low-rank tuning transformations, channel-wise affine modulation, and bounded dispersion adjustments constrain adaptation. At deployment, the small target-day set is encoded once to jointly adapt neural tuning and dispersion, with no target-day generator fine-tuning.

Empirical support under a limited target-day budget. With 50 real target-day trials and 50 generated trials, BURST achieves a PER of 47.96%, compared with 47.02% using 100 real trials, a difference of 0.94 percentage points (Table 6). These results support the use of a small target-day set to adapt the generative mapping under the conditions evaluated.


We hope these clarifications address your concerns,We would appreciate if you could consider raising your score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers for their constructive feedback and engagement with our work. We are particularly grateful to Reviewers Y3EB and gxey for recognizing the practical importance of few-shot cross-day calibration and the value of synthetic data generation in reducing patient burden.


Recap of Core Contributions

Our work addresses cross-day neuroprosthesis calibration through three technical contributions that together answer why adaptive dispersion is needed, how the generator adapts, and what structure is preserved:

1. Neural Recording Perspective (Why): We identify that recorded spike counts exhibit sub-Poisson variability due to refractoriness and threshold-detection dead time, with the degree of suppression varying across days. This motivates adaptive dispersion modeling conditioned on target-day pilot data to capture the day-specific variability structure that changes under cross-day drift.

2. Algorithmic Innovation (How): We introduce a birth-death count bridge with simulation-free analytic reverse-rate targets (Lemma 2), enabling explicit rate-dependent dispersion control through Conway-Maxwell-Poisson formulation. The generator-first paradigm places synthesis before calibration, directly addressing the constraint that every real trial imposes patient burden whereas synthetic trials are cost-free once conditioned.

3. Sequential Label Generation (What): We perform the first sequence-level neural synthesis for speech decoding, preserving phoneme-to-neural temporal structure including coarticulation and boundary dynamics through alignment-free phoneme conditioning. Prior work conditions only on single categorical labels, which cannot capture the sequential structure essential for speech decoding.


Clarifications on Common Concerns

Performance (PER)

Some reviewers (Reviewer zTqq) raised concerns about performance. We clarify that our reported numbers already match state-of-the-art baselines:

  • Table 2 (long-horizon protocol): All-real FT achieves approximately 31% PER, consistent with the 30% suggested by Reviewer zTqq (Ryoo et al. (2025)).
  • Table 1 (medium-horizon protocol): All-real FT achieves approximately 42% PER due to reduced pretraining data (7 days versus 12 days).

Our testing protocol follows realistic deployment scenarios: we train only on historical data and test on future target days, never allowing the model to peek at target-day neural distributions during pretraining. BURST achieves its gains while reducing calibration burden (from several hundred sentences requiring hours of recording time to 50-100 sentences).

SBP Generation

We discuss SBP-based fine-tuning in the appendix. Our conclusion is that SBP features provide limited additional value for cross-day transfer because they do not capture the day-specific count dispersion structure that changes across days, the primary cross-day challenge addressed by our count-bridge.

We have additionally conducted experiments on the long-horizon protocol to demonstrate this point:

Table: SBP generation on the long-horizon protocol, N=100N=100, PER (mean across Days 15-21).

MethodAvg. PERΔ\Delta vs. NR-alone
NR-alone (100 real).361-
BURST (NR+NG, count-only).337-.024
BURST (NR+NG) + SBP.329-.032

Key findings: SBP generation improves PER by .008 (.337 → .329), modest compared to the count-based adaptation benefit (.361 → .337, Δ=.024). The SBP contribution primarily reflects decoder baseline improvement with spectral features rather than enhanced cross-day transfer capability. Our count-bridge already captures the day-specific sub-Poisson variability structure; SBP provides incremental benefit through complementary frequency-domain information but does not address the core cross-day transfer challenge.


We have provided detailed responses to each reviewer addressing specific concerns, additional experiments where appropriate, and clarifications on existing analyses in the manuscript. We hope these responses demonstrate the technical soundness and practical value of our approach. Given the clarifications on performance alignment with SOTA baselines, the technical novelty of our contributions, and the practical significance of calibration efficiency, we would appreciate if the reviewers could consider raising their scores.

Area_Chair_viDNOpenReview ↗

Dear Reviewers

Since the author-reviewer discussion ends soon, I would like to encourage the reviewer to review the author's rebuttals and communicate with them about whether the author's responses are satisfactory and any remaining concerns.

Thanks for your time and effort in reviewing for ICLR 2027.

Best wishes,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.