- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies few-trial calibration for cross-session neural decoding. The authors propose DiDA, which represents session-specific drift using a low-dimensional Gaussian distribution and generates alignment transformations from sampled drift states. The model is pretrained on Day 0 data and augmented versions of those data. During adaptation to a new session, only the mean and standard deviation of the drift distribution are updated, while the shared encoder, latent dynamics, and behavioral decoder remain fixed. The authors evaluate supervised, unsupervised, and preference-guided adaptation on macaque center-out and random-target reaching tasks. Under the evaluated few-trial, trial-by-trial adaptation settings, DiDA generally outperforms the compared methods.
Critical strengths
- The paper addresses the limited availability of calibration data at the beginning of a new session, with a clear motivation and supporting experiments. DiDA performs well in few-trial sequential adaptation, particularly in the unsupervised setting. These results support constraining adaptation through a pretrained structure and are relevant to reducing repeated calibration in BCIs.
- Preference-guided adaptation is a direction worth exploring. The drift distribution produces candidate decoded trajectories, and preferences between these trajectories are used to update the distribution. This offers an alternative when precise behavioral labels are unavailable. The authors clearly acknowledge that the current preferences are generated by an oracle using ground-truth trajectories. Although these experiments do not establish effectiveness with human feedback, they provide an initial proof of concept.
Critical weaknesses
-
The respective contributions of the low-dimensional constraint and the probabilistic representation remain unclear. The appendix examines regularization terms, dynamics constraints, and drift dimensionality, but does not directly compare learning a drift distribution, optimizing a deterministic low-dimensional state, and optimizing an unconstrained transformation matrix under controlled conditions. The role of the distribution in candidate sampling and preference optimization is relatively clear. For supervised and unsupervised adaptation, however, it remains difficult to determine how much of the improvement comes from the low-dimensional constraint, probabilistic modeling, or other pretraining choices. This limits my assessment of the contribution of each design choice.
-
Several aspects of the baseline comparisons and the consistency of the gains need clarification. In Table 2, some baselines have lower mean performance under supervised adaptation than under unsupervised adaptation, along with substantial variability across sessions. This does not by itself indicate inadequate tuning, but warrants an explanation in terms of hyperparameters, loss weights, and update budgets. The differences between the present setting and related methods such as FDA [1] and SPINT [2], and the reasons for not comparing them directly, also deserve discussion. Moreover, the aggregate means and standard deviations do not show whether DiDA improves performance in most sessions or how sensitive the results are to random seeds. These issues limit my confidence in the comparative conclusions.
[1] Wang, P., Qi, Y., Wang, Y., and Pan, G. Flow Matching for Few-Trial Neural Adaptation with Stable Latent Dynamics. In Proceedings of the 42nd International Conference on Machine Learning, Proceedings of Machine Learning Research, 267:64059–64083, 2025. [2] Le, T., Fang, H., Li, J., Nguyen, T., Mi, L., Orsborn, A. L., Sümbül, U., and Shlizerman, E. SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding. In Advances in Neural Information Processing Systems, 38, 2025.
Questions for authors
- Could you compare three adaptation variants using the same pretrained DiDA model: (i) the original probabilistic drift distribution; (ii) a deterministic low-dimensional drift vector, retaining the fixed transformation-generating network; and (iii) an unconstrained session-specific transformation matrix optimized directly, without the low-dimensional parameterization? This would help distinguish the effects of probabilistic modeling and the low-dimensional constraint. If the existing Align-seqVAE experiment already provides the third control, please clarify whether the shared backbone, pretraining data, and adaptation settings are the same, and identify the corresponding results.
- The appendix reports DiDA’s adaptation steps and examines the update interval. Could you clarify how the hyperparameters, loss weights, and update budgets were selected for each main baseline? Is there any further diagnosis or explanation for the lower means and greater variability after adding behavioral supervision in some Table 2 results, such as CycleGAN with 25 adaptation trials? Please also discuss the comparability of FDA [1], CRRL [2], and SPINT [3] to the present setting and the reasons for not including direct comparisons. If comparable results are already available, please provide them.
- The reported means and standard deviations are computed across target sessions. Were the main comparisons also repeated with multiple random seeds? Could you provide per-session paired performance differences between DiDA and the main baselines, to show whether the average gains reflect improvements in most sessions or are driven primarily by a small number of sessions?
5 comments
W2 & Q2 Additional baselines
We thank the reviewer for these suggestions. We have added FDA and SPINT as additional baselines under our sequential adaptation protocol.
For SPINT, its gradient-free adaptation avoids test-time parameter updates, but also limits its ability to explicitly correct session-specific drift using incoming trials. In contrast, DiDA progressively refines its drift distribution as more observations become available.
For FDA, we selected its MMD variant as the baseline because it demonstrated better performance in the original paper. However, we observed substantial sensitivity to hyperparameter selection and occasional representation collapse. A key issue is that both source and target distributions are produced by the same trainable encoder during adaptation. Consequently, minimizing MMD allows both distributions to shift together toward degenerate solutions without necessarily preserving behaviorally relevant representations. We therefore consider FDA a less reliable baseline for evaluating strictly unsupervised adaptation, as its reported performance may benefit from information that would not be available in practical label-free settings.
We will include the additional baseline results and implementation details in the revised manuscript.
Tab. 3 Comparison of DiDA with FDA and SPINT on the CO task.
| Supervision Mode | Method | 10 | 25 | 50 | Offline |
|---|---|---|---|---|---|
| Supervised | DiDA | 0.6961 0.1027 | 0.7646 0.0698 | 0.7968 0.0539 | 0.8263 0.0414 |
| Supervised | FDA | 0.5400 0.1063 | 0.6736 0.0877 | 0.7720 0.0629 | 0.8939 0.0197 |
| Unsupervised | DiDA | 0.5262 0.1740 | 0.6107 0.1414 | 0.6637 0.1002 | 0.6660 0.0973 |
| Unsupervised | FDA | 0.2052 0.1457 | 0.0952 0.0680 | 0.0271 0.0234 | 0.3139 0.2024 |
| Unsupervised | SPINT | 0.2508 0.2290 | 0.2648 0.2079 | 0.2852 0.2255 | 0.3390 0.2151 |
Tab. 4 Comparison of DiDA with FDA and SPINT on the RT task.
| Supervision Mode | Method | 10 | 25 | 50 | Offline |
|---|---|---|---|---|---|
| Supervised | DiDA | 0.4923 0.1642 | 0.5302 0.1734 | 0.5798 0.0970 | 0.6067 0.0803 |
| Supervised | FDA | 0.2470 0.1249 | 0.2888 0.1441 | 0.2774 0.1689 | 0.5901 0.0727 |
| Unsupervised | DiDA | 0.1909 0.5471 | 0.3331 0.3159 | 0.3625 0.3009 | 0.4537 0.1295 |
| Unsupervised | FDA | 0.1959 0.0913 | 0.1038 0.0611 | 0.0111 0.0308 | 0.2922 0.1395 |
| Unsupervised | SPINT | 0.0543 0.1548 | 0.1556 0.1278 | 0.2072 0.1413 | 0.2511 0.1428 |
W2 & Q4 Multi-seed evaluation, paired per-session differences, and performance variability.
We thank the reviewer for this suggestion. We have repeated the main experiments with multiple random seeds and will update the reported results by averaging across seeds. We have also added paired per-session performance differences between DiDA and the main baselines to assess whether the improvements are consistent across sessions rather than driven by a few favorable cases. The per-session analysis helps explain the large variability and occasional performance degradation observed in Table 2, particularly for CycleGAN under supervised adaptation with 25 trials. We found that a small number of challenging target sessions exhibit unusually poor decoding performance, substantially lowering the mean and increasing the standard deviation. Thus, these fluctuations are largely driven by extreme session-specific outcomes rather than a consistent degradation across all sessions. We will include the multi-seed results and per-session analyses in the revised manuscript to make these effects more transparent.
Importantly, our comparison focuses on adaptation efficiency under limited sequential observations, rather than maximum performance with unrestricted optimization. The strong offline performance of the baselines, together with the relatively comparable supervised results, indicates that they are capable of learning effective transformations. However, their performance degrades more substantially under unsupervised sequential adaptation, where limited observations provide weaker and noisier gradient signals. This interpretation is further supported by Appendix F.1, where Align-seqVAE improves substantially with larger update intervals, while DiDA remains relatively stable. These findings suggest that the performance differences are closely related to the methods' sensitivity to sparse sequential updates, rather than simply inadequate hyperparameter tuning. We will clarify these details in the appendix.
Tab. 5 Per-session supervised adaptation performance () on the CO task.
| Day offset | DiDA | Align-SeqVAE | ERDiff | CycleGAN | FDA |
|---|---|---|---|---|---|
| 1 | 0.8670 0.0072 | 0.6401 0.0519 | 0.6163 0.0844 | 0.5139 0.0544 | 0.5620 0.106 |
| 2 | 0.8231 0.0158 | 0.7381 0.0152 | 0.6183 0.1080 | 0.6226 0.0307 | 0.7136 0.0112 |
| 3 | 0.6234 0.0607 | 0.4551 0.0350 | 0.4897 0.0557 | 0.4069 0.0449 | 0.4954 0.1055 |
| 9 | 0.7886 0.0076 | 0.6987 0.0277 | 0.6951 0.0060 | 0.6437 0.0180 | 0.5857 0.0328 |
| 10 | 0.7543 0.0449 | 0.7322 0.0274 | 0.6155 0.1204 | 0.6478 0.0209 | 0.6722 0.0388 |
| 14 | 0.7893 0.0278 | 0.7459 0.0179 | 0.6897 0.0523 | 0.6976 0.0100 | 0.7182 0.0357 |
| 15 | 0.8195 0.0108 | 0.7608 0.0203 | 0.7608 0.0372 | 0.6804 0.0323 | 0.7447 0.0070 |
| 16 | 0.6711 0.0959 | 0.7602 0.0167 | 0.7158 0.0549 | 0.6706 0.0514 | 0.7313 0.0217 |
| 36 | 0.7524 0.0124 | 0.7302 0.0529 | 0.6797 0.0294 | 0.6255 0.0239 | 0.6780 0.0222 |
| 37 | 0.7320 0.1027 | 0.8239 0.0172 | 0.6975 0.0371 | 0.6835 0.0284 | 0.7623 0.0157 |
| 38 | 0.7903 0.0543 | 0.7734 0.0171 | 0.6086 0.0812 | 0.6824 0.0182 | 0.7455 0.0436 |
We thank the reviewer for the insightful feedback and constructive recommendations. We will address these points in the revision to further improve the quality and completeness of our work.
W1 & Q1 Probabilistic vs. deterministic and unconstrained adaptation
The third variant is already covered by Align-seqVAE, which optimizes a full session-specific transformation matrix using the same backbone architecture, pretraining data, and adaptation protocol. We have additionally evaluated the second variant by replacing the probabilistic drift distribution with a deterministic low-dimensional vector while keeping the pretrained DiDA model fixed. The deterministic variant performs substantially worse, particularly under unsupervised adaptation. We attribute this difference to the ambiguity of drift estimation under limited supervision. While deterministic adaptation optimizes a single drift state that may be sensitive to noisy or weakly constrained alignment signals, probabilistic adaptation optimizes the expected loss over sampled drift states, introducing a smoothing effect on the optimization landscape. This can reduce sensitivity to sharp local minima and encourage solutions that perform consistently across neighboring drift states. Together, these comparisons highlight the complementary benefits of probabilistic modeling and structured low-dimensional parameterization.
Tab. 1 Comparison of probabilistic, deterministic, and unconstrained drift adaptation on CO.
Tab. 2 Comparison of probabilistic, deterministic, and unconstrained drift adaptation on CO.