acceptodds
Under review as a conference paper at ICLR 2027

SAUNA: Source-Anchored Unsupervised Neural Adaptation for Cross-Subject EEG Decoding

Abstract

Decoding seen images from EEG has advanced rapidly, but much of that progress has been measured subjects, with a decoder trained and evaluated on the same person. However, practical adoption will require a more challenging cross-subject setting, in which the decoder must generalize to a person whose labeled data was never observed. We ask what actually changes when a decoder trained on some brains is applied to an unseen brain. A contrastive decoder aligns two representation spaces, an image side and a signal side, and we answer this question for each side by direct measurement. On the image side, we show that the depth at which a vision backbone should supply its alignment target is a property of the visual hierarchy rather than of the individual. The alignment target can therefore be : estimated once from source subjects and shared by every new subject. Implemented as per-dimension selection over the depth of a multi-backbone feature pool, the anchored target yields a zero-shot decoder that attains \textbf{37.72\\%} 200-way top-1 accuracy on THINGS-EEG2, surpassing the strongest published result of 24.0%. On the signal side, we show that about half of the attainable accuracy is subject-specific and beyond the reach of source data alone. To recover this component without labels, we introduce SAUNA (Source-Anchored Unsupervised Neural Adaptation), which uses only EEG from the target subject and never observes the trial-to-image pairing. SAUNA reads three properties of the subject's signal from such data: its position relative to the image manifold, its distribution shape, and its trial noise. One objective is instantiated for each property and optimized jointly with the anchor, so that the representation adapts without drifting from image space. SAUNA raises accuracy from 37.25% for a matched control to \textbf{47.77\\%}, a gain that is positive for all ten held-out subjects, recovers roughly a third of the subject-specific gap, and is, to our knowledge, the strongest label-free cross-subject result reported to date.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG decoding, visual neural decoding, cross-subject generalization, unsupervised domain adaptation

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_xpu8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies EEG-to-image retrieval for a new subject without using that subject's EEG–image pairings. It first uses source-subject EEG to select a visual feature layer for each dimension across five frozen backbones. It then trains an EEG encoder with source contrastive supervision and three target objectives: information maximization over the training image gallery, moment matching, and contrastive consistency between two halves of repeated responses. The target stimulus set and repetition groups are known, but image identities are withheld. On THINGS-EEG2, the source-only model reaches 37.72% Top-1 accuracy, and SAUNA improves a matched control from 37.25% to 47.77%. Testing uses a 200-image gallery and EEG averaged over 80 repetitions per image.

Critical strengths
  • The adaptation gain is substantial. Top-1 increases by 10.52 percentage points and Top-5 by 9.02 points against the matched control. All ten subjects improve, and the paired Top-1 gains are 11.80, 10.50, and 9.25 points across the three seeds (Table 1).

  • The calibration-budget experiment is useful. Figure 4b reports both the benefit of additional target data and the failure at very small budgets. The two-minute setting reduces accuracy for all ten subjects, identifying a practical limitation at small calibration budgets.

  • The authors disclose several relevant protocol details. The authors explain the target-subject MVNN preprocessing, the history of feature-pool selection, and differences between published baselines. Reporting the final epoch throughout also avoids choosing checkpoints by their test accuracy.

Critical weaknesses

Weakness 1: The added value of the new components is unclear

Most methodological components have close precedents: intermediate-layer supervision, multi-backbone fusion, InfoMax, and covariance alignment are established approaches. Du et al. (2026) also report that preferred visual depths are broadly consistent across subjects. The main methodological additions appear to be per-dimension depth selection and the particular combination of target objectives. Neither addition is tested against the most informative simple control. Table 5 gives 36.4% for a fixed intermediate depth, compared with 37.72% for per-dimension selection, but these results use one and three seeds respectively. In Table 1, InfoMax alone contributes 7.18 of the full 10.52-point gain. The missing InfoMax + CORAL and InfoMax + repetition comparisons would show whether all three terms are needed. The full gain exceeds the sum of the individual gains by only 0.09 points, which is too small to support the claim of synergy without an uncertainty estimate.

Weakness 2: The conclusions extend beyond the conditions tested

All main experiments use ten subjects from THINGS-EEG2. There is no second dataset or cross-session evaluation to test whether the depth preferences and adaptation gains persist under different recording conditions. The method also assumes that target calibration stimuli come from the known training image set and that repeated responses are correctly grouped. Neither assumption is varied in the experiments. It remains unclear whether adaptation retains its benefit under incomplete gallery coverage or repetition-group errors. These tests would directly examine the conditions required by InfoMax and repetition consistency. Without them, the practical claims should be limited to the controlled calibration setting used here.

Weakness 3: The headline comparisons leave important differences uncontrolled

The authors state that the external methods were not rerun and that their preprocessing, electrode sets, and checkpoint selection differ. Although these differences are disclosed, they prevent Table 2 from isolating the proposed method's contribution from protocol differences. The strongest evidence remains the internal matched control. A comparison with a simpler intermediate-layer target and standard adaptation losses under the same setup would be more informative about the contribution. The already-cited SAMGA reports 34.4% cross-subject Top-1, but is absent from Table 2. SAUNA's 37.72% is higher, although the setups differ. This is useful context for the emphasis on the 24.0% baseline; I do not count the absence of an arXiv-only comparison as a reason for rejection. Similarly, the 14.8% SATTC result uses a different encoder and access protocol, so it cannot establish the relative merits of the two adaptation procedures, as the authors acknowledge.

Weakness 4: Table 6 does not explain the variance accounting

For CLIP ViT-B/32, the reported components are 47.6% for shape, 4.1% for subject offset, and 0.7% for interaction. These sum to 52.4%. The other rows likewise account for only about half of the variance. The remaining variance, any residual term, and the denominator used for each percentage are unspecified. The decomposition equation is needed to interpret these values. There is a related ambiguity: Table 6 calls the subject component an “offset” but describes it as a “scale,” while Figure 1b divides each profile by its subject mean. These are different operations. Until their relationship is specified, I cannot assess the interpretation of the small interaction term. Even if that calculation is correct, similar average depth profiles do not establish that subject-specific selection of individual dimensions has no benefit, as the paper suggests.

Weakness 5: The post-hoc experiment changes more than the source anchor

Post-hoc adaptation starts from a trained source model, removes the source loss, and freezes the image mapping. Full SAUNA instead uses the target objectives from the start and trains the image mapping jointly with source supervision. The resulting 40.90% versus 47.77% comparison therefore confounds source supervision, image-mapping updates, and adaptation timing. The lower post-hoc score could reflect the missing source loss, the frozen image mapping, or when adaptation begins. To test the anchor's role, both runs should start from the same checkpoint, use the same number of updates, and treat the image mapping identically, with only the source loss switched on or off. The comparison supports the complete joint-training procedure, but does not isolate the contribution of the source anchor.

Weakness 6: The acquisition and training costs limit the deployment claim

The 47.77% Top-1 result is obtained after averaging 80 EEG repetitions for each test image. Even with that averaging, more than half of the queries are not retrieved correctly at rank one. Despite the substantial benchmark improvement, evaluation with fewer repetitions is needed to support the claims about interactive use. Calibration also takes about an hour to obtain most of the gain, and each new subject requires joint training from scratch with the source data. The paper does not report training time, peak memory, or the end-to-end calibration latency for a new subject. With these costs unmeasured, it is difficult to assess the practical benefit of discarding target image pairings while still requiring a known stimulus set and repetition groups.

Questions for authors

See Weaknesses.

1 comment

I have not seen a response from the authors. My concerns about the methodological contribution and experimental controls remain unresolved, so I am maintaining my original rating.

Reviewer_C6KXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies cross-subject EEG to image retrieval on THINGS-EEG2. It makes two claims. First, the best depth to read a vision backbone is shared across subjects. This is shown by a variance decomposition of ridge encoding R2 over subject and depth. So the image-side target is fixed from source subjects, using per-dimension layer selection over five frozen backbones. This gives 37.72% zero-shot 200-way top-1. Second, about half of the accuracy is subject specific. SAUNA adapts to a new subject with unlabeled training-split EEG. It adds three losses to the supervised source loss. \mathcal{L}{\mathrm{im}} is entropy minimization with a diversity term over the training-image gallery. \mathcal{L}{\mathrm{coral}} matches the mean and covariance of target and source embeddings. \mathcal{L}_{\mathrm{rep}} is a contrastive loss between two halves of a trial’s repetitions. SAUNA raises top-1 from 37.25% for a matched control to 47.77%. The paper also reports an oracle at 73.53%, a calibration time analysis, a subject-count scaling curve, and notes on protocol differences across baselines.

Critical strengths

The reporting is honest in places. Each adaptation result has a matched control and a paired per-subject sign test. The appendix measures noise floors and shows that best-epoch numbers on this benchmark are inflated. This is useful for the field. The paper lists protocol differences across published baselines. These are electrode set, epoch selection on test data and preprocessing. The scoring rule analysis shows that assignment scoring adds close to 20 points. Readers should know this.

Critical weaknesses

The label-free claim is weaker than stated. The adaptation step sees the target’s full training-split EEG, the known training image set, and the grouping of repetitions. The diversity term in \mathcal{L}{\mathrm{im}} assumes each gallery image appears equally often. That is information about the label marginal. The grouping tells the model which trials share an image. This is set-level supervision. The paper admits the setting is easier than zero-shot. But it still calls 47.77% the strongest label-free cross-subject result. The only comparison is SATTC at 14.8%, which was not re-run and uses a different access model. No standard adaptation baseline is run under the same access. Euclidean alignment, SHOT, DANN and plain fine-tuning are all missing. There is also no baseline with a few minutes of labeled target data. Without these, the reader cannot tell if the three losses matter. Most of the gain comes from \mathcal{L}{\mathrm{im}}, which is SHOT-style entropy minimization. CORAL is also existing work. The paper says it does not offer a new primitive. So the contribution rests on the comparison, and the comparison is missing. The zero-shot headline is not a controlled result. The 37.72% is compared to numbers copied from papers with different encoders, preprocessing and epoch selection. The paper admits this. The feature pool of five backbones was chosen by within-subject accuracy on this benchmark. Table 5 shows that reading every backbone at fixed relative depth 0.4 gives 36.4%. Per-dimension selection gives 37.72%. So the proposed anchoring adds about 1.3 points over a fixed mid-depth, on one seed. The rest is the known intermediate-layer effect from Benchetrit, Du and Zheng. On 17 channels the paper reaches 34.37% against 21.8% and 20.0% for the closest prior work. This large gap is not explained. No ablation covers pool composition, number of backbones, or selection granularity. No published method is re-run on this pool. The diagnostic claims are not tested. Depth invariance is shown with ridge R2 on ten subjects from one dataset. No experiment lets depth vary per subject and checks decoding. The conclusion that per-subject depth has no headroom is untested. The paper says about half of the accuracy is beyond the reach of source data. Table 4 shows +11.1 top-1 per doubling of source subjects, with no flattening at nine. That contradicts the claim. The oracle also trains on the target’s labeled data, so the 35.8 point gap mixes several effects. The statement that the three losses are complementary rests on 10.52 against a sum of 10.43. That is noise. Configurations and loss weights were screened on subjects 01, 03, 06 and 08. These subjects are also in the reported ten. The text says nothing was tuned on reported test numbers. These two statements conflict. The control baseline appears as 37.15, 37.25, 37.35 and 37.40 in different sections. The depth-selected row is 37.72.

Questions for authors

What do Euclidean alignment, SHOT, DANN and entropy minimization alone give under your exact access model? What does fine-tuning with 10, 30 and 60 minutes of labeled target data give? What is the result for fixed depth 0.4, per-backbone depth and per-dimension depth, over three seeds? Which backbones can be removed? If depth is chosen per subject with an oracle, how much does decoding change? Please redo the screening on subjects that are excluded from the final report. Do the loss weights and batch size change the result? What do ATM, NICE and one recent baseline score when trained on your feature pool and your preprocessing? How do you reconcile the Table 4 scaling curve with the claim that the subject gap cannot be closed with source data? Does \mathcal{L}_{\mathrm{im}} still help if the target subject sees only part of the training image set, or if image frequencies are unequal?

1 comment

The author has provided no further response, I am maintaining my score.

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

SAUNA anchors EEG-to-visual decoding on labeled source subjects and adapts to a held-out subject using only unlabeled training EEG. It separates zero-shot leave-one-subject-out retrieval from label-free target adaptation and reports 200-way top-1/top-5 accuracy on ten THINGS-EEG2 subjects.

Critical strengths
  1. The evaluation explicitly separates zero-shot (37.72% top-1) from target-adapted (47.77%) results and excludes target test EEG from adaptation, making the protocol and source of improvement clear.

  2. Table 1 gives a matched 37.25 +/- 0.65 control versus 47.77 +/- 0.92 for SAUNA, with improvement for all ten subjects and three-seed reporting.

Critical weaknesses
  1. The principal results come from ten subjects on one EEG dataset. Transfer to different acquisition settings or datasets remains unestablished, limiting the scope of the cross-subject generalization claim.

  2. InfoMax alone adds 7.18 points against the matched control, versus 10.52 for the full method. Without leave-one-objective-out joint ablations, the contributions of the other objectives under combined training remain unclear.

Questions for authors
  1. Can the authors provide leave-one-objective-out ablations, especially the full objective without InfoMax, to clarify the contributions under joint training?

  2. Is there existing evidence that the fixed source anchor and feature pool transfer to an independent EEG acquisition setting? A new external-dataset study would be a useful extension.

1 comment

Having read the other reviews, I am maintaining my rating of 2, as my concerns about the contributions of individual objectives under joint training and generalization beyond THINGS-EEG2 remain unresolved.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.