acceptodds
Under review as a conference paper at ICLR 2027

FunctionalMap: Learning Data-Driven Coordinates for Cross-Subject Transformer Modeling of Heterogeneous Intracranial Recordings

21 views

Abstract

Aggregating intracranial recordings across subjects is challenging because electrode count, placement, and regional coverage vary widely. Standard anatomical coordinates such as MNI provide a shared reference frame, but often fail to capture true functional similarity: even at matched anatomical coordinates, the targeted brain region and underlying neural dynamics can differ substantially between individuals. We propose FunctionalMap, a scalable representation-learning framework that learns a data-driven functional coordinate system for intracranial neural recordings. An encoder trained with contrastive objectives maps short local field potential segments into a subject-agnostic embedding space, inducing electrode identities that are locality-sensitive to region-specific neural signatures rather than anatomical location. These functional coordinates are then used as tokens for a transformer that models inter-regional relationships across variable channel sets. We evaluate this framework on a 20-subject basal ganglia–thalamic dataset collected during flexible rest/movement recording sessions with heterogeneous electrode layouts. The learned functional space forms clear, region-consistent clusters, supports within-subject discrimination, and transfers zero-shot to unseen channels. In cross-region LFP prediction, a single transformer using functional coordinates outperforms models using MNI coordinates or region labels, without subject-specific heads or fine-tuning. Performance improves with more training subjects; functional coordinates also improve spectral band-power prediction, and the masked-region result replicates with separately trained models on a public human cohort. On public mouse visual-behavior recordings, a trained functional encoder supports transformer modeling of pooled hit-versus-miss behavior prediction in unseen animals (zero-shot) and improves upon anatomical coordinates.

Primary area:
applications to neuroscience & cognitive science
Keywords:
intracranial neural recordings, cross-subject learning, representation learning, contrastive learning, neural time series, transformer models

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

This papaer introduce a framework called FunctionalMap and it is designed to model brain activity across different individuals, even when dealing with diverse or messy intracranial recordings. Instead of relying on traditional anatomical locations, it learns functional coordinates from neural signals. To address variations in electrode placement, channel count, and regional coverage across subjects, the authors train a CNN-based encoder with region-supervised contrastive learning to map short LFP segments into a shared 32-dimensional embedding space. These embeddings act as channel identifiers within a Transformer architecture tailored for cross-region neural signal prediction. The authors mainly evaluate the method on multi-region LFP recordings from 20 human subjects. The learned embeddings show region-specific clustering and can generalize to held-out channels. When used in the Transformer, they perform better than MNI coordinates and region labels for LFP reconstruction. The pooled model also outperforms several subject-specific baselines, including FIR, TCN, and GRU. The authors further test the method on spectral band-power prediction, a public human dataset, and mouse Neuropixels recordings for behavioral prediction. Overall, the paper explores whether learned functional coordinates can help combine neural recordings from different subjects, especially when their recording setups are quite different.

Critical strengths
  1. The paper considers cross-subject modeling of intracranial recordings with varying electrode placements, channel counts, and regional coverage. The proposed framework allows recordings from different subjects and sessions to be processed by a shared Transformer without subject-specific heads or fine-tuning. This provides a practical approach to aggregating heterogeneous neural recordings.
  2. The authors compare functional coordinates with MNI coordinates and region labels under the same Transformer architecture. The experiments show improvements in cross-region LFP prediction, with additional capacity-matched comparisons and replication on a separate public human dataset. These results provide empirical support for the usefulness of learned functional coordinates in the evaluated prediction tasks.
Critical weaknesses
  1. The paper mainly compares the proposed functional coordinates with MNI coordinates and region labels. Other baseline methods, such as FIR, TCN, and GRU, are only trained on individual subjects. However, there are already some related works, including Population Transformer [1], seegnificant [2], and BaRISTA [3], which also focus on modeling neural recordings across subjects with different electrode configurations. Although the authors mentioned these works in the Related Work section, they did not provide direct experimental comparisons. I understand that these methods may use different datasets or tasks, but I think comparisons with at least some of them would help better show the advantages of FunctionalMap. Also, the proposed method mainly combines existing contrastive learning and Transformer architectures. Without comparisons with these related methods, it is difficult to understand what the main methodological improvement is compared with previous studies.
  2. Another concern is about what the learned functional coordinates actually represent. The functional encoder is trained using region-supervised contrastive learning, which directly encourages signals from the same brain region to be close in the embedding space. Therefore, I am not fully convinced that the observed clustering results can show that the model has learned functional information beyond region identity or some basic spectral features. I think adding comparisons with simple spectral features or other self-supervised embeddings would make this claim more convincing. In addition, in Appendix A.15, some subjects are excluded when training the functional encoder, but the downstream Transformer is still trained using data from all subjects. Therefore, this experiment cannot fully demonstrate zero-shot generalization to completely unseen subjects. Since cross-subject modeling is one of the main motivations of this paper, I think a more strict leave-one-subject-out evaluation for the whole framework would be necessary.

References [1] Chau, G., et al. (2025). Population Transformer: Learning Population-level Representations of Neural Activity. ICLR 2025.
[2] Mentzelopoulos, G., et al. (2024). Neural Decoding from Stereotactic EEG: Accounting for Electrode Variability Across Subjects. NeurIPS 2024.
[3] Oganesian, L., Hashemi, S., & Shanechi, M. M. (2025). BaRISTA: Brain Scale Informed Spatiotemporal Representation of Human Intracranial Neural Activity. arXiv preprint.

Questions for authors
  1. I noticed that the paper discusses BaRISTA, Population Transformer, and seegnificant in Section 2, but none of them is included in the experiments. Since these methods also address cross-subject neural modeling with different electrode configurations, why were they not considered as baselines? I think comparing with at least some of them would help clarify the contribution of FunctionalMap.
  2. I find the learned functional coordinates interesting, but I wonder what they actually represent. Since the encoder is trained with region labels, could the embeddings mainly capture region-specific spectral features? Also, the coordinates are extracted from each channel's own signals, while MNI coordinates and region labels do not have access to this information. Would simple PSD or band-power features from the same 10-second segment achieve similar results? This seems like an important control.
  3. In Appendix A.15, the test subjects are excluded from encoder training, but the Transformer is still trained on all 20 subjects. So this seems to show encoder generalization rather than generalization of the full framework. Have the authors considered a LOSO experiment where both models are trained without the test subject?
  4. In Table 1, Pooled-FUNC performs much better than Single-FUNC. Could this improvement mainly come from having more training data? A pooled TCN or GRU baseline would help clarify this.
  5. How stable are the learned functional coordinates? For example, would different 10-second segments from the same electrode give similar embeddings? I am also curious whether they remain stable across rest, movement, or different recording sessions.
  6. Could the authors clarify the referencing scheme used for the micro-contact recordings? I wonder whether shared-reference effects or common neural activity could partly explain the cross-region prediction results. Would alternative referencing or temporal shifting change the results?
10 comments
AuthorsOpenReview ↗

Addressing Q4: Pooled TCN Baseline

We thank the reviewer for this important suggestion. We agree that the increased training data could contribute to the improvement of our pooled Functional Transformer. Indeed, enabling effective aggregation across subjects is a central motivation of our work. However, aggregating intracranial recordings is challenging because subjects have different numbers of electrodes without one-to-one channel correspondence.

Conventional multivariate TCNs and GRUs assume fixed input/output channel identities, making direct multi-subject training difficult without architectural modifications. Our Transformer instead uses channel-wise tokenization and attention to model interactions between arbitrary sets of recording channels. This design is consistent with recent intracranial modeling approaches such as seegnificant (Mentzelopoulos et al., 2024) and BaRISTA (Oganesian et al., 2025), which also employ Transformer-based architectures.

New experiment: Pooled TCN baseline

To directly address this concern, we implemented a channel-agnostic TCN and trained one shared model across the pooled 20-subject dataset.

Architecture: Each source electrode is processed independently using a shared five-block residual TCN (64 hidden channels, kernel size 5, dilations 1, 2, 4, 8, 16). Its features are conditioned on the frozen 32-D MSC functional coordinates and combined using masked mean aggregation, allowing variable numbers of source electrodes. A two-block TCN decoder, conditioned on each target electrode's functional coordinates, reconstructs the withheld VO waveforms. The model has 311,385 trainable parameters and uses no subject-specific heads or subject IDs.

Training and evaluation: We preserved the original preprocessing, 1-second windows with 500-ms overlap, chronological 70/15/15 splits, 25-ms patches, and masked-region prediction task. We used the same objective, MSE+0.05(1−r)\mathrm{MSE}+0.05(1-r), with AdamW, batch size 16, learning rates of 2×10−42\times10^{-4} for the backbone and 3×10−43\times10^{-4} for the output head, and early stopping. Evaluation uses subject-level Fisher-zz-averaged test Pearson correlation, as in Table 1.

Results (mean ± SEM):

ModelLeft (n=18n=18)Right (n=20n=20)
Single-subject TCN (Table 1)0.250 ± 0.0570.279 ± 0.048
Pooled TCN + Functional0.320 ± 0.0480.368 ± 0.046
Pooled Functional Transformer0.481 ± 0.0430.475 ± 0.045

The pooled TCN improves over the original subject-specific TCN baseline in both hemispheres. However, despite using pooled recordings and the same functional coordinate system, it remains below the pooled Functional Transformer (Δr=0.162\Delta r=0.162 left and 0.1080.108 right).

These results support two conclusions. First, non-Transformer architectures can also benefit from pooled recordings when adapted to heterogeneous electrode configurations. Second, access to more training data alone does not reproduce the performance of our Functional Transformer. Its ability to preserve individual channel representations and model their interactions provides an important architectural advantage over this pooled TCN implementation.

AuthorsOpenReview ↗

Addressing Q1: Functional coordinates improve BaRISTA on the same cross-region task

Thank you for suggesting a comparison with an existing multi-subject model. We adapted BaRISTA (Oganesian et al., 2025) to reconstruct the masked VO region on the same dataset and chronological splits used in our main experiment.

We retained BaRISTA’s dilated CNN tokenizer, 250-ms patches, and 12-layer encoder with 64-dimensional tokens, four attention heads, and 10% dropout. We used 1-second windows at 1 kHz to match our original inputs and outputs, masked all VO channels, and trained directly on waveform reconstruction. The three coordinate variants used the same backbone, with about 0.83 million trainable parameters each.

CoordinatesLeft (N=18N=18), Pearson rr (mean ± SD)Right (N=20N=20), Pearson rr (mean ± SD)
Region-MLP320.166±0.1890.166 \pm 0.1890.156±0.1620.156 \pm 0.162
MNI-MLP320.220±0.2010.220 \pm 0.2010.262±0.1770.262 \pm 0.177
MSC Functional320.291±0.1840.291 \pm 0.1840.311±0.2070.311 \pm 0.207

Functional coordinates improved reconstruction over region labels in 17 of 18 subjects on the left (mean paired gain 0.1250.125, Holm-adjusted p=0.00069p=0.00069) and all 20 subjects on the right (gain 0.1560.156, Holm-adjusted p=0.00034p=0.00034). Compared with MNI, functional coordinates improved reconstruction in 15 of 18 subjects on the left (gain 0.0720.072, Holm-adjusted p=0.0238p=0.0238) and 13 of 20 on the right (gain 0.0490.049, Holm-adjusted p=0.104p=0.104). The gain over MNI was statistically significant only on the left. Although the adapted BaRISTA backbone scored lower overall than our original Transformer, both benefited from functional coordinates. Functional coordinates achieved the highest mean reconstruction score in both hemispheres with the adapted BaRISTA backbone, showing that the benefit is not specific to our Transformer.

The main contribution of FunctionalMap is a way to pool heterogeneous recordings across subjects using shared, signal-derived channel coordinates. This is particularly useful for our 50 µm microcontacts, where nearby contacts can have similar estimated MNI coordinates despite recording different signals. Our other evaluations also use small recording sites, specifically microwire LFPs from the public human movie-watching dataset and Neuropixels recordings from the Allen Visual Behavior dataset. For reference BaRISTA was originally trained and evaluated on Brain Treebank recordings from 0.8-mm sEEG contacts.

AuthorsOpenReview ↗

Addressing Q5: Functional coordinate stability across time, contacts, and movement

Thank you for your comment. We ran two new analyses using the frozen FunctionalMap encoder. In both, we report median cosine distances (1−cos⁡(θ)1-\cos(\theta)) across subjects, with 95% subject-bootstrap confidence intervals. Lower values indicate more similar coordinates.

1. Stability over time and across contacts

For each contact, we estimated its coordinate from the first 10 seconds, exactly as in the Transformer. We compared it with coordinates from up to eight later, non-overlapping 10-second clips, spanning up to about 10 minutes. We matched the time offsets across all three comparisons.

ComparisonLeft (n=20n=20)Right (n=20n=20)
Same contact, later clip0.224 [0.178, 0.264]0.296 [0.191, 0.322]
Different contact, same region0.512 [0.411, 0.700]0.593 [0.534, 0.680]
Different contact, different region0.706 [0.624, 0.833]0.642 [0.573, 0.875]

In all 20 subjects in each hemisphere, later clips were closer to the reference from the same contact than to time-matched clips from another contact in the same region. At the group level, different contacts in the same region were closer than contacts in different regions. This shows that the coordinates retain reproducible differences between contacts within a region, rather than mapping every contact in a region to the same location.

We also matched channel labels across separate recording sessions within each subject. Compared with the within-period baseline on the same clips, distances increased from 0.273 to 0.401 on the left and from 0.329 to 0.464 on the right. This suggests greater drift across periods, although matching labels alone does not verify physical contact identity. For context Functional coordinates are also calculated per session (first 10 seconds).

2. Stability across EMG-defined rest and movement

These clinical recordings did not follow a uniform behavioral protocol. We focused on right-hand movement, a common recorded activity. We averaged four right-hand EMG envelopes and smoothed the result over 200 ms. A 10-second clip was labeled movement if at least half its samples exceeded the recording-period median, and rest otherwise. In the sampled clips, the median fraction above threshold was about 77% for movement and 18% for rest. We compared clips from the same contact and recording period at matched time gaps across all three pair types.

Same-contact comparisonLeft, contralateral (n=19n=19)Right, ipsilateral (n=19n=19)
Rest–rest0.260 [0.216, 0.294]0.273 [0.185, 0.319]
Movement–movement0.316 [0.223, 0.429]0.272 [0.198, 0.356]
Rest–movement0.287 [0.253, 0.341]0.329 [0.262, 0.417]
Paired excess across states−0.004 [−0.021, 0.042]+0.037 [0.006, 0.070]

The paired excess is the rest–movement distance minus the average of the two within-state distances(behavioral state-dependent increase in coordinate distance), computed before aggregating across subjects. Across 20 recording periods, we found no detectable excess on the contralateral side and a small increase on the ipsilateral side. Together, the two analyses show that a coordinate estimated from a single 10-second clip retains contact-specific information over time, while allowing for some variation across behavioral states and recording periods.

Reviewer_EYmqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper a two-stage representation learning framework for modeling intracranial EEG recordings across subjects. First, a CNN encoder learns functional embeddings for recorded channels through a contrastive learning objective that pulls channels from same regions close to each other in the embeddings space across subjects. These embeddings are then used as positional embeddings to encode functional/spatial information in the Functional Transformer, which is trained using a self-supervised objective to reconstruct masked regions from other simultaneously recorded regions. The model is evaluated on region decoding, neural waveform and spectral prediction on a 20-subject clinical cohort. The model is also separately trained and evaluated on a human movie-watching dataset and a mice Neuropixel LFP dataset using in similar evaluation settings.

Critical strengths
  • The paper addresses an important problem: whether learning data-driven channel coordinates can improve representation learning for intracranial EEG despite variability in recording and electrode placement across subjects.

  • Functional coordinates improve cross-region reconstruction over MNI coordinates and region labels in both the 20-subject cohort and the public movie-viewing dataset, providing evidence that functional coordinates learned from data capture shared channel information beyond anatomical and region identity (MNI coordinates and region information) that can be unreliable.

Critical weaknesses

Overall, the functional embeddings could provide a useful contribution to intracranial EEG modeling if the following concerns are addressed, and I would update my score accordingly.

  • The usefulness of the functional embeddings for human downstream decoding is not evaluated. Most of the analysis in the main text focuses on region classification and cross-region reconstruction, which are not sufficient to establish representation quality for intracranial EEG decoding tasks. Evaluating human tasks, such as those in Brain Treebank, would help establish whether the additional information provided by functional embeddings improves task-relevant decoding over MNI coordinates or region labels. Relatedly, I find the simulations and region decoding results less informative than the missing downstream decoding evaluation, particularly since the authors themselves argue that anatomical labels are unreliable, making region classification alone a limited evaluation metric.

  • There is no direct comparison with existing multi-subject intracranial models such as BaRISTA or Brant, which also support neural reconstruction. The standard prediction baselines are useful, but they do not establish whether FunctionalMap improves over existing state-of-the-art models. Comparing cross-region reconstruction with these models, or comparing them on a common downstream decoding task, would help contextualize the contribution: whether the proposed functional coordinates offer an advantage over representations already learned by existing methods.

  • Moreover, the functional embeddings are trained separately on each dataset. It is not evaluated whether these embeddings can transfer across datasets or enable training across datasets with different recorded brain regions. This limits the evidence that the proposed coordinates can support larger-scale integration of heterogeneous intracranial recordings.

Questions for authors
  • Considering that the Functional transformer is trained on all subjects in Table A.10, the subjects are held out only from the training for functional embeddings. can this evaluation be extended to completely held-out subjects, also excluding them from transformer training?
  • As authors mention, the electrode coordinates or region information can be unreliable. Is supervision by regions necessary for learning useful functional embeddings? The waveforms already contain spectral and other channel-specific information which I believe are being learned in the CNN functional encoder. How does the method compare with embeddings learned from the same waveforms without region supervision, or with spectral features?
  • Are the embeddings really subject-invariant? Omitting subject IDs does not guarantee this, since the waveforms themselves may contain subject-specific information. Can the authors evaluate whether subject identity is predictable from either the functional embeddings or from the learned representations of the transformer?
  • PSC and MSC are compared only on region decoding. Does MSC actually improve cross-region reconstruction over PSC?
9 comments
AuthorsOpenReview ↗

Direct Comparison of PSC and MSC in Downstream Cross-Region Prediction

Thank you for this important question. We agree that better held-out-channel classification alone does not establish that MSC performs better in downstream reconstruction. We therefore directly compared PSC and MSC functional coordinates in the cross-region LFP prediction Transformer on our full 20-subject dataset.

Experimental setup. We trained separate Transformers using PSC-derived and MSC-derived 32-D functional coordinates. All other settings were identical, including Transformer architecture (1,366,555 trainable parameters), source/target channels, chronological train/validation/test splits, input windows, optimization, and evaluation. Only the functional encoder used to obtain each electrode's fixed coordinate was changed.

Results. We report subject-level Fisher-averaged Pearson correlation (mean ±\pm SD). For completeness, we include the anatomical and capacity-matched controls already reported in Appendix A.11.

CoordinateLeft (n=18n=18)Right (n=20n=20)
Region0.2373 ±\pm 0.20740.2505 ±\pm 0.1823
Region-MLP320.2416 ±\pm 0.20230.2553 ±\pm 0.1838
MNI0.3183 ±\pm 0.20310.2561 ±\pm 0.1734
MNI-MLP320.3186 ±\pm 0.20710.3413 ±\pm 0.1837
PSC-Functional0.4340 ±\pm 0.17850.4545 ±\pm 0.1893
MSC-Functional0.4817 ±\pm 0.18440.4759 ±\pm 0.1999

MSC significantly outperformed PSC in both hemispheres:

  • Left: Mean paired improvement Δr=0.0477\Delta r = 0.0477, Holm-corrected p=4.35×10−4p = 4.35 \times 10^{-4}.
  • Right: Mean paired improvement Δr=0.0214\Delta r = 0.0214, Holm-corrected p=6.39×10−3p = 6.39 \times 10^{-3}.

Significance was assessed using paired two-sided Wilcoxon signed-rank tests on subject-level scores, with Holm correction across the five PSC comparisons within each hemisphere.

Importantly, PSC also significantly outperformed both capacity-matched anatomical and region-based controls in both hemispheres (all Holm-corrected p≤1.07×10−3p \leq 1.07 \times 10^{-3}).

These findings provide two conclusions. First, the stronger held-out-channel performance of MSC translates into significantly better downstream cross-region reconstruction, supporting our choice of MSC for the main experiments. Second, PSC also improves substantially over anatomical coordinates, showing that the benefit of learned functional coordinates is not limited to one contrastive objective.

We appreciate the reviewer highlighting this missing comparison and will add these results and the corresponding protocol to the revised manuscript.

AuthorsOpenReview ↗

Addressing critical weakness: Functional coordinates improve BaRISTA on the same cross-region task

Thank you for suggesting a comparison with an existing multi-subject intracranial model. Our main experiments held the Transformer architecture fixed to isolate what the data-driven coordinates contribute when pooling recordings across subjects. We agree that it is also important to test whether this advantage carries over to an architecture used in prior work. To do so, we adapted BaRISTA (Oganesian et al., 2025) to reconstruct the masked VO region using the same recordings and chronological splits as in our main experiment.

We retained BaRISTA’s dilated CNN tokenizer, 250-ms patches, and 12-layer encoder with 64-dimensional tokens, four attention heads, and 10% dropout. We used 1-second windows at 1 kHz to match our original inputs and outputs, masked all VO channels, and trained directly on waveform reconstruction. The three coordinate variants used the same backbone, with about 0.83 million trainable parameters each.

CoordinatesLeft (N=18N=18), Pearson rr (mean ± SD)Right (N=20N=20), Pearson rr (mean ± SD)
Region-MLP320.166±0.1890.166 \pm 0.1890.156±0.1620.156 \pm 0.162
MNI-MLP320.220±0.2010.220 \pm 0.2010.262±0.1770.262 \pm 0.177
MSC Functional320.291±0.1840.291 \pm 0.1840.311±0.2070.311 \pm 0.207

Within the BaRISTA backbone, functional coordinates improved reconstruction over MNI in 15 of 18 subjects on the left (mean paired gain 0.0720.072, Holm-adjusted p=0.0238p=0.0238) and 13 of 20 on the right (gain 0.0490.049, Holm-adjusted p=0.104p=0.104). The gain over MNI was statistically significant only on the left. Compared with region labels, functional coordinates improved reconstruction in 17 of 18 subjects on the left (gain 0.1250.125, Holm-adjusted p=0.00069p=0.00069) and all 20 subjects on the right (gain 0.1560.156, Holm-adjusted p=0.00034p=0.00034). Although the adapted BaRISTA backbone scored lower overall than our original Transformer, both benefited from functional coordinates.

Functional coordinates yielded the highest mean reconstruction scores in both hemispheres, showing that shared, signal-derived channel coordinates can improve pooled reconstruction in this setting without relying on our particular Transformer.

BaRISTA was originally trained and evaluated on Brain Treebank recordings from 0.8-mm sEEG contacts. Our cohort uses 50 µm microcontacts, where nearby contacts can have similar estimated MNI coordinates despite recording different signals.

AuthorsOpenReview ↗

Are the embeddings really subject-invariant?

Thank you for pointing out that omitting subject IDs does not guarantee subject invariance. We tested how well a regularized linear classifier could predict subject identity from the frozen Functional V2 coordinates. We fit a separate probe for each region and hemisphere, so differences in which regions were recorded could not drive identification. The probe trained on multiple 10-second clips from other recording periods and was tested on the first 10-second embedding of each held-out contact, the same coordinate used by the Transformer.

Functional V2 across 20 subjects. The table shows balanced top-1 accuracy. Uniform-guess accuracy was 1/201/20 (5%).

RegionHemisphereHeld-out periodHeld-out period + channel labels
GPiLeft16.8%14.8%
GPiRight14.8%14.1%
STNLeft18.7%14.0%
STNRight20.7%13.1%

Subject identity was partly decodable, but the probe misidentified the subject on roughly 85–87% of predictions in the stricter split. We would not necessarily expect exactly 5% accuracy. Even in a shared coordinate system, subjects can differ in which parts of a region their electrodes sample, whether those differences are described by anatomical locations or by functional coordinates. We were inspired to also answer the same question on the MNI coordinates, thus performed the following experiment as well.

Anatomical control in GPi. To see how much subject information could be recovered from location alone, we repeated the probe with MNI coordinates. We held out entire recording periods and electrode location groups, and verified that no test XYZ appeared in training for the same subject. Both representations used identical splits. We report GPi because 19 of 20 subjects per hemisphere had enough distinct locations for this comparison.

HemisphereMNIFunctional V2
Left GPi25.0%12.9%
Right GPi16.9%12.2%

Interestingly, subject ID was numerically less predictable from Functional V2 than from MNI in both hemispheres. This control illustrates that even a shared anatomical coordinate system can carry subject information. The probe measures what can be read out from the coordinates, not whether the Transformer uses that information.

Reviewer_h8BCOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MNI coordinates attempt to put the electrode locations of different subjects into a common space. This is often not perfect, since the same MNI location in different subjects may have different functionality, even if only due to subject variation. This paper presents a method for learning position embeddings based on contrastive learning. The contrastive learning pulls together representations of activity from the same region. The evaluation is a cross-region prediction task, where data from one region is given, and data from another region must be predicted.

Critical strengths
  • The paper is well written, and the method is explained clearly.

  • The resulting representations cluster well semantically

Critical weaknesses
  • Can you explain more clearly why the method is not circular? As I understand it, you want to avoid using an atlas like MNI to give position embeddings. But some atlas is needed to produce the region labels for the contrastive learning. Doesn’t this introduce the same problems?

  • For the sake of exposition, please introduce method 2 (line 196) with a sentence or two of motivation.

  • Line 225: It would also be good to test on a general forecasting/classification task. It seems like the biggest thing missing is a demonstration that these embeddings can be a drop-in replacement for MNI for one of the papers you cite. The biggest doubt that remains for me is the thought that existing models can incorporate this region information already in the course of training, which would mean that MNI coordinates are sufficient and this process is redundant.

Questions for authors
  • Please number equations.

  • Line 75: Is this a typo?

  • Line 203: the sum over aa includes other positives. How do you sample batches to avoid the case where the batch includes a balance of negatives and positives?

  • Figure 2B: mentions beta bursts, but this is nowhere else mentioned in the main text? Can you elaborate on its significance?

6 comments
AuthorsOpenReview ↗

Replacing MNI with functional coordinates in BaRISTA

Thank you for asking whether functional coordinates can serve as a drop-in replacement for MNI in an existing model. We tested this directly by adapting BaRISTA (Oganesian et al., 2025) to our masked-VO reconstruction task and comparing MNI, region labels, and functional coordinates within the same backbone.

We retained BaRISTA’s dilated CNN tokenizer, 250-ms patches, and 12-layer encoder with 64-dimensional tokens, four attention heads, and 10% dropout. To match our main experiment, we used the same recordings, chronological splits, and 1-second windows at 1 kHz. We masked all VO channels and trained each variant from scratch to reconstruct their waveforms. The three variants used about 0.83 million trainable parameters each.

CoordinatesLeft (N=18N=18), Pearson rr (mean ± SD)Right (N=20N=20), Pearson rr (mean ± SD)
Region-MLP320.166±0.1890.166 \pm 0.1890.156±0.1620.156 \pm 0.162
MNI-MLP320.220±0.2010.220 \pm 0.2010.262±0.1770.262 \pm 0.177
MSC Functional320.291±0.1840.291 \pm 0.1840.311±0.2070.311 \pm 0.207

Replacing MNI with functional coordinates improved reconstruction in 15 of 18 subjects on the left (mean paired gain 0.0720.072, Holm-adjusted p=0.0238p=0.0238) and 13 of 20 on the right (gain 0.0490.049, Holm-adjusted p=0.104p=0.104). The gain over MNI was statistically significant only on the left. Compared with region labels, functional coordinates improved reconstruction in 17 of 18 subjects on the left (gain 0.1250.125, Holm-adjusted p=0.00069p=0.00069) and all 20 subjects on the right (gain 0.1560.156, Holm-adjusted p=0.00034p=0.00034).

Functional coordinates yielded the highest mean reconstruction scores in both hemispheres, with significant gains over region labels on both sides and over MNI on the left (the right-side gain over MNI was 0.0490.049, Holm-adjusted p=0.104p=0.104). The benefit is therefore not specific to the Transformer used in our main experiment.

AuthorsOpenReview ↗

Addressing circularity in region-supervised functional coordinates

Anatomy guides training, but the signal determines the coordinate. In our human cohort, experts assigned coarse region labels from co-registered MRI and CT images. Those labels can still be imperfect. They tell the encoder which recordings should be similar during training, but they do not assign a fixed coordinate to each electrode. The learned features help distinguish regions, but the coordinate itself is data-driven. It depends on the content of the recorded LFP, not just the electrode's physical location. Two contacts in the same region can therefore have different coordinates, and a contact's coordinate can change across sessions if its signal changes. Similarly, a bad quality recording not demonstrating region specific signatures would not have coordinates close to that region,

The approach generalizes to new recordings with no anatomical data. In Appendix A.15, we held out four patients per fold from encoder training. Reconstruction performance was similar whether or not a patient was included in encoder training. The predictor still trained on all patients, so this tests generalization of the encoder and the capability of the transformer to train on subjects with no anatomical data available. In the mouse experiment, the test animals were excluded from both training stages. Functional coordinates achieved a mean AUROC of 0.7250.725, versus 0.7120.712 with anatomical coordinates and 0.6850.685 with region labels.

Computing a new coordinate does not require anatomical localization. In our human cohort, anatomical localization used MRI and CT. In the Allen mouse dataset, fine-grained localization relied on tracing probe tracks in post-mortem images. Once trained, the functional encoder can assign a coordinate from a short LFP segment without waiting for either localization step.

The benefit extends beyond region classification. Region clustering alone would not resolve the circularity concern. In masked-region LFP prediction, functional coordinates improved mean reconstruction from 0.2420.242 to 0.4820.482 on the left and from 0.2550.255 to 0.4760.476 on the right relative to the capacity-matched region-label model (Appendix A.11). The gain is in predicting held-out neural activity, not in recovering the labels used to train the encoder.

AuthorsOpenReview ↗

Motivation for the modified supervised contrastive objective

Thank you for the suggestion. We will add the following motivation before introducing Method 2: “PSC scores each preselected pair independently. We also tested a batch-based objective that lets each LFP segment be compared with multiple same-region and different-region examples at once, to see whether this helps the encoder learn features that distinguish regions across subjects and sessions.”

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.