acceptodds
Withdrawn from ICLR 2027

Repeat-Calibrated Neural Representations for Cross-Subject Generalization in Stereoelectroencephalography

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Model–brain alignment within recorded individuals does not establish transfer across people with different electrode layouts or across auditory domains. We introduce the Repeated Auditory Stereoelectroencephalography Dataset (RASD) and Repeat-Calibrated Neural Fields (RCNF) to investigate how response reliability and spatial aggregation affect such generalization. RASD records 30 pediatric and adolescent participants listening to one story and six music excerpts, each scheduled three times, linking 2,052 signal-valid contacts to stimulus events and providing 2,050 matched contact coordinates. RCNF uses training-repeat consistency to calibrate contacts before coordinate-based, participant-balanced aggregation, constructing a fixed-dimensional representation for a shared predictor. Four tasks evaluate held-participant prediction on new repetitions, reduced source coverage, and bidirectional story–music transfer; other participants supply neural responses to the test content. Across 21 participants and 161 auditory contacts, calibration improves all four endpoints over uniform reliability weighting after correction for 16 comparisons, with absolute gains of .00069–.00128. Temporal history improves three tasks; the full spatial representation has no corrected advantage over either spatial control. All aligned endpoints exceed time-shifted references. Together, RASD and RCNF make repeat reliability an explicit component of neural representation design, providing a data and method foundation for predictors that can be reused across people and auditory materials.

Primary area:
datasets and benchmarks
Keywords:
Stereoelectroencephalography (SEEG), Cross-subject neural prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_5rhzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces RASD, a dataset of repeated auditory SEEG recordings from 30 pediatric and adolescent participants, and RCNF, a method for predicting a held-out participant's neural responses from other participants' responses to the same stimuli. RCNF combines repeat-reliability weighting, spatial pooling, equal averaging across source participants, and temporal history with a shared Bayesian-ridge predictor.

The benchmark uses 161 auditory contacts from 21 participants and tests repeat generalization, story–music transfer, and reduced source-electrode coverage. Reliability weighting produces small improvements across four endpoints, while spatial complexity has no corrected advantage over simpler pooling. Prediction requires source neural recordings of the test content; held-participant training-domain recordings are also used for eligibility and scaling.

Critical strengths
  • Useful repeated-stimulus dataset design. Shared story and music excerpts, repeated presentations, and electrode coordinates allow response reliability, electrode coverage, and cross-participant transfer to be studied together within one acquisition protocol. This is a potentially valuable resource for evaluating methods across heterogeneous electrode layouts, provided the release includes adequate documentation and quality-control information.
  • Additional controls support a limited reliability finding. Appendix E reports that true contact–weight assignments outperform shuffled assignments in the earlier story fold (AUC gain approximately .00172; positive in 20/21 participants). This supports the relevance of which contacts receive the weights, although it is exploratory and does not replace a comparison with learned pooling.
Critical weaknesses
  • The benefit over alternative methods and its practical importance remain unclear. The current benchmark compares variants of RCNF, without a learned-pooling alternative. Reliability gains are approximately 0.00069–0.00128 across the four endpoints. Appendix E does provide within-participant repeat correlations, which are useful context, but explicitly notes that they are neither permission-matched predictors nor theoretical ceilings. The practical importance of the incremental gains therefore remains unresolved; low raw correlations alone do not establish poor performance on noisy responses.
  • Acquisition provenance and physiological quality control remain incomplete. If playback hardware, acoustic-contamination checks, and treatment of epileptiform activity remain unspecified. These could affect how we interpret the decoding results.
Questions for authors
  1. Model comparison: Have the authors compared RCNF with lightweight learned pooling, such as Deep Sets or target-conditioned attention, under the same participant splits and calibration budget? An adapted Population Transformer would be another relevant comparison. Does the prescribed weighting offer an advantage over learning how to combine electrodes?
  2. Acquisition and quality-control clarification: Were stimuli delivered through earphones/headphones or loudspeakers, and at what levels? Could the authors specify the referencing and notch-filter settings? Beyond the stated numerical eligibility rule, how were “signal-valid” contacts identified, and were bad channels, epileptiform activity, and acoustic/equipment artifacts screened? If artifacts were found, how were they handled while preserving neural responses? Acoustic contamination is documented in other intracranial modalities (Roussel et al., 2020); this motivates checking the setup, not assuming these SEEG data are contaminated.
  3. Effect-size interpretation: Appendix E provides useful within-participant repeat references and explains why they are not theoretical ceilings. Can the authors relate these references more directly to the primary task results, or provide another justified estimate of predictable signal with uncertainty, to help interpret the small calibration gains?
  4. Dataset contribution: How does RASD complement earlier repeated natural-sound intracranial datasets, particularly Norman-Haignere et al. (2022), and shared speech/music datasets such as Berezutskaya et al. (2022)? Which aspects of its acquisition design or released data enable evaluations that those resources do not?
Reviewer_FPBiOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the Repeated Auditory Stereoelectroencephalography Dataset (RASD) and Repeat-Calibrated Neural Fields (RCNF) to study cross-subject and cross-domain generalization in auditory SEEG responses. RASD includes 30 pediatric and adolescent participants listening to one story and six music excerpts, each repeated three times, with 2,050 matched contact coordinates. Results show calibration improves all four endpoints over uniform reliability weighting after multiple comparison correction.

Critical strengths

The paper addresses a practical issue by jointly testing cross-subject transfer, repeat generalization, and cross-domain transfer within a single dataset design.

Critical weaknesses

The reported performance gains from calibration are extremely small in absolute terms , and the full spatial representation offers no corrected advantage over simpler spatial controls, raising serious doubts about the practical significance and robustness of the proposed method.

The evaluation cohort is limited to a clinically sampled pediatric and adolescent group with short auditory materials and uneven contact coverage, and the inference is explicitly nominal because the cohort informed development, undermining the strength of the statistical claims.

Questions for authors

Given that the absolute gains from calibration are below 0.0013 and the full spatial representation shows no corrected advantage over global pooling, what concrete practical scenario justifies the added complexity of RCNF over simpler baselines?

How do you ensure that the participant-balanced aggregation does not simply dilute informative dense sampling rather than genuinely improving cross-subject transfer, especially when coverage varies from 2 to 20 contacts per participant?

Since the cohort informed development and outer folds share training and test-source participants, what specific steps would be needed to validate these findings on a truly independent dataset before the method could be considered reliable for reuse?

The channel-level repeat consistency analysis shows that training-repeat consistency does not reliably predict held-repeat consistency in the selected example. How does this observation reconcile with the claim that repeat reliability is a useful calibration signal?

What is the theoretical or empirical justification for the fixed soft mapping in the calibration weight formula, and how sensitive are the results to the chosen clipping bounds and the zero-correlation baseline of 0.5?

Given that temporal history helps only three of four tasks and coordinate-dependent features show no corrected advantage, what evidence supports the continued use of multiscale spatial kernels rather than a simpler fixed-scale or global pooling design?

Can you provide more detail on the exact procedure for selecting the single scale in the SingleScale variant, and how does that selection interact with the reported lack of advantage for the full multiscale representation?

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

This paper introduces a dataset (RASD) and a predictor (RCNF) with SEEG recording as input and high-gamma neural response as predicted output for "unseen" participants. For the same auditory stimulus, RCNF uses inputs of other participants and their coordinates in MNI space and a query coordinate and predicts a time signal for this query.

RCNF uses the contact repeat reliability for spatial weights around the target location. It generalizes within domain (same stimulus) as well as cross-domain (music to story and vice versa).

Critical strengths
  • Cross-subject tranfer without correspondence of SEEG layouts is an open challenge.

  • The dataset is very carefully designed.

  • The paper clearly separates the methodological contributions of reliability weighting, coordinate pooling, temporal history.

  • Uses appropriate controls for interpretability

  • Authors clearly acknowledge limitations

Critical weaknesses
  • I am not familiar with the numbers in the literature but reliability gains around 0.001 when the absolute values are also low raise questions about usability. Please correct me with references in the literature.

  • Ablations over local vs global scales hardly show any difference.

  • There is no comparison to any other set-based or graph-based method.

  • A clipped affine transformation of pairwise repeat correlation may be sensitive to short windows and correlations with SNR levels.

Questions for authors
  • We need a comparison with at least one more method. This is the most significant drawback of the paper. Very open for acceptance if we get a calibration with respect to baselines.

  • How should I interpret the low gains (see weakness)?

  • Explain the transformation in eq. (2).

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

Related papers

Discussion (0)

Sign in to comment.